# VAAS: Vision-Attention Anomaly Scoring for image manipulation detection in digital forensics

## Full Text

This Markdown is an automated text extraction for search and reading tools. The [hosted PDF](VisionAttentionAnomolyScoringImageManipulationDetection.pdf) is the authoritative publication source.

Source PDF: [VisionAttentionAnomolyScoringImageManipulationDetection.pdf](VisionAttentionAnomolyScoringImageManipulationDetection.pdf)
Source SHA-256: `7579332d4742cf037c3b92d294fe45cb3b6cc004ee9df99a79bff9dadec9ddd4`
Generated with `pdftotext -nopgbrk`; formatting and reading order may differ from the PDF.

---

Forensic Science International: Digital Investigation 56 (2026) 302063

Contents lists available at ScienceDirect

Forensic Science International: Digital Investigation
journal homepage: www.elsevier.com/locate/fsidi

DFRWS EU 2026 - Selected Papers from the 13th Annual Digital Forensics Research Conference Europe

VAAS: Vision-Attention Anomaly Scoring for image manipulation detection
in digital forensics
Opeyemi Bamigbade a, Mark Scanlon b,*, John Sheppard a
a
b

Forensics and Security Research Group, South East Technological University, Waterford, Ireland
Forensics and Security Research Group, School of Computer Science, University College Dublin, Dublin, Ireland

A R T I C L E I N F O

A B S T R A C T

Keywords:
Digital forensics
Image manipulation detection
Tamper localisation
Explainable AI
Vision transformers
Segmentation
Attention mechanisms
Anomaly scoring

Recent advances in AI-driven image generation have introduced new challenges for verifying the authenticity of
digital evidence in forensic investigations. Modern generative models can produce visually consistent forgeries
that evade traditional detectors based on pixel or compression artefacts. Most existing approaches also lack an
explicit measure of anomaly intensity, which limits their ability to quantify the severity of manipulation. This
paper introduces VISION-ATTENTION ANOMALY SCORING (VAAS), a novel dual-module framework that integrates global
attention-based anomaly estimation using Vision Transformers (ViT) with patch-level self-consistency scoring
derived from segmentation embeddings. The hybrid formulation provides a continuous and interpretable
anomaly score that reflects both the location and degree of manipulation. Evaluations on the DF2023 and CASIA
v2.0 datasets demonstrate that VAAS achieves competitive F1 and IoU performance, while enhancing visual
explainability through attention-guided anomaly maps. The framework bridges quantitative detection with
human-understandable reasoning, supporting transparent and reliable image integrity assessment. The source
code for all experiments and corresponding materials for reproducing the results are available open source.

1. Introduction
The reliability of digital images as admissible evidence has become a
critical concern in modern forensic investigations. Multimedia forensic
techniques can play a central role in verifying the authenticity of visual
content presented in legal, journalistic, and intelligence contexts
(Spichiger and Adelstein, 2025). In these contexts, one of the primary
objectives is the detection and localisation of image manipulations,
ensuring that digital evidence remains trustworthy in court. In the 2024
digital forensic practitioner survey (Hargreaves et al., 2024), ~20 % of
practitioners report they encounter deepfakes “occasionally” or “often”
in their investigations.
Recent advances in generative models such as GANs and diffusion
networks have significantly increased the difficulty of detecting
tampered or fabricated images, as these models produce semantically
coherent and artefact-free edits that evade traditional forensic cues.
Traditional image manipulation techniques are broadly classified into
three categories: (1) splicing, where content from one image is copied
into another; (2) copy-move, which involves duplicating a region within
the same image; and (3) inpainting, which fills missing regions with

synthetic or reconstructed content (Wu et al., 2019; Verdoliva, 2020; Ma
et al., 2023). Examples of these manipulations are shown in Fig. 1. Early
forensic methods relied on identifying pixel-level inconsistencies,
compression artefacts, or metadata discrepancies. However, these cues
are often removed or concealed in AI-enhanced manipulations that
employ diffusion models or adversarial synthesis pipelines.
With the increasing accessibility of generative AI tools, images can
now be modified with remarkable realism and semantic coherence (Ma
et al., 2023). These models adjust fine-grained features, such as texture,
illumination, and object boundaries, leaving minimal forensic traces. As
a result, conventional forensic algorithms focused solely on low-level
features struggle to capture the spatial and contextual relationships
that define visual authenticity. Addressing this challenge requires
methods capable of learning high-level spatial dependencies and
detecting subtle inconsistencies in attention and feature representation.
In this context, attention mechanisms and Vision Transformers
(ViTs) provide a promising foundation for forensic analysis. Their ability
to capture global context and model long-range dependencies makes
them suitable for identifying inconsistencies between authentic and
manipulated regions. Building on this principle, this paper introduces a

* Corresponding author.
E-mail addresses: opeyemi.bamigbade@postgrad.setu.ie (O. Bamigbade), mark.scanlon@ucd.ie (M. Scanlon), john.sheppard@setu.ie (J. Sheppard).
https://doi.org/10.1016/j.fsidi.2026.302063
Available online 24 March 2026
2666-2817/© 2026 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/bync-nd/4.0/).

O. Bamigbade et al.

Forensic Science International: Digital Investigation 56 (2026) 302063

splicing detection using boundary or texture irregularities (Bappy et al.,
2019) provided interpretable results suitable for forensic reporting.
However, these handcrafted cues degrade when images are recom­
pressed, filtered, or regenerated through AI pipelines, limiting their
robustness.
2.2. AI-generated manipulations and modern challenges
Generative models such as GANs and diffusion networks can modify
lighting, structure, and scene semantics while preserving pixel-level
coherence (Ma et al., 2023). Such manipulations lack the artefacts
exploited by traditional detectors and often evade signal-based methods,
especially when trained on large-scale synthetic datasets. This shift has
created a need for forensic approaches that reason about spatial coher­
ence rather than relying solely on low-level noise patterns.

Fig. 1. Example of content-changed-based image manipulation techniques
(copy–move, splicing, and removal). Extracted from Shi et al. (2024).

2.3. Deep learning in image forensics

hybrid framework that couples global attention analysis with local
manipulation segmentation for interpretable anomaly detection.

CNN-based forensic models improve robustness by learning
discriminative features directly from data (Nguyen et al., 2022; Zhuang
et al., 2021). More recent transformer-based systems leverage attention
mechanisms to capture long-range dependencies and semantic in­
consistencies, making them effective for generative manipulations (Hao
et al., 2021; Wang et al., 2022). Despite strong performance, most deep
models act as black boxes and do not provide an interpretable, contin­
uous measure of tampering intensity. This gap motivates hybrid ap­
proaches that combine global context, local consistency, and
interpretable scoring.

1.1. Contributions of this work
The proposed VAAS framework introduces a novel hybrid anomalyscoring formulation that integrates global attention deviations with
local self-consistency, offering a unified and interpretable measure of
image manipulation that has not been addressed in the literature. The
key contributions are summarised as follows:
• Dual-module forensic architecture: A unified design integrating a
Vision Transformer-based Forensic Attention Extractor (Fx) for global
anomaly representation and a segmentation-based Manipulation
Segmentor (Px) for localised tampering analysis.
• Hybrid anomaly scoring mechanism: A principled fusion strategy
that combines global attention deviation and local spatial inconsis­
tency into a single hybrid score quantifying image integrity.
• Explainable forensic interpretability: The generation of attentiondriven anomaly maps and segmentation masks that provide visual
and interpretable evidence, thereby improving transparency and
trust in forensic decision-making.
• Empirical validation and benchmarking: A comprehensive eval­
uation on two datasets demonstrates competitive image manipula­
tion detection, stable localisation, and strong generalisation across
traditional (CASIA v2.0) and generative manipulation types
(DF2023).

2.4. Motivation for VAAS
The proposed VISION-ATTENTION ANOMALY SCORING (VAAS) framework
builds on these developments by combining global attention analysis
with local self-consistency scoring. VAAS introduces a hybrid mecha­
nism that quantifies the likelihood of manipulation while producing
visually interpretable evidence maps. In doing so, it bridges the gap
between quantitative performance and qualitative transparency, align­
ing computational detection with the practical requirements of digital
forensic validation.
3. Related work
The continuous advancement in artificial intelligence, particularly in
computer vision, has driven substantial progress in image forgery
detection. Traditional methods for digital image integrity verification,
which rely on pixel-level anomaly detection, compression artefacts, or
metadata analysis, are increasingly inadequate against the complex
manipulations produced by modern generative models. This section
reviews major research directions in image manipulation detection,
authenticity verification, and the emerging role of attention and crossattention mechanisms in forensic image analysis.

All experimental code and reproducibility resources are available
open source at https://github.com/OBA-Research/VAAS-experiments.
2. Background
The authenticity of digital images remains a central concern in
forensic investigations, where visual evidence must be traceable,
reproducible, and defensible. Traditional forensic methods focused on
low-level signal artefacts, while modern manipulation techniques
increasingly rely on generative models that alter global semantics
without leaving obvious pixel-level traces. This section summarises the
evolution of image manipulation detection, highlighting why global­
–local reasoning is needed for contemporary forensic analysis.

3.1. Image manipulation detection
AI-driven image manipulation detection has become an essential
field across journalism, law enforcement, and national security, where
visual authenticity is critical (Singh, 2022). The introduction of GANs
and diffusion models has enabled seamless and photorealistic forgeries
(Guarnera et al., 2024). Earlier forensic approaches relied on
hand-crafted cues such as pixel-level inconsistencies, colour filter array
artefacts, or JPEG quantisation traces. However, these methods fail
against artefact-free content generated by advanced synthesis pipelines
(Du et al., 2020).
The rise of deep learning, particularly CNNs, marks a shift toward
automated representation learning for forensic detection. CNN-based
architectures, such as XceptionNet and its variants, which are widely

2.1. Classical forensic approaches
Early forensic techniques relied on analysing pixel-level disturbances
introduced by editing workflows. Examples include DCT-block in­
consistencies, CFA artefact disruption, and resampling traces (Monika
et al., 2021; Mahdian and Saic, 2008). Copy–move detection using
SIFT/SURF descriptors (Pandey et al., 2014; Christlein et al., 2012) and
2

O. Bamigbade et al.

Forensic Science International: Digital Investigation 56 (2026) 302063

applied in the DeepFake Detection Challenge and FaceForensics++
benchmark (Nguyen et al., 2022; Zhuang et al., 2021), have improved
generalisation across manipulation domains. Matern et al. (2019)
leveraged semantic inconsistencies, such as missing reflections or
asymmetric facial cues, while Li et al. (2018, 2020) explored spatial
artefacts arising from face warping and resampling. Feng et al. (2020)
introduced a two-stage triplet-loss framework that learns discriminative
embeddings for real and fake faces. Likewise, Kingra et al. (2024) in­
troduces SFormer, an end-to-end transformer architecture that in­
tegrates spatial and temporal information to enhance the accuracy of
deepfake detection. Despite their effectiveness, CNN-based systems
primarily capture local texture information and lack the capacity to
model long-range spatial dependencies. This is an essential factor in
detecting complex manipulations.

The evolution of forensic datasets in image manipulation detection
reflects a shift toward greater diversity, realism, and semantic richness.
Modern benchmarks such as DF2023 embody this transition by
combining traditional tampering operations with AI-driven generative
manipulations. By evaluating VAAS across CASIA and DF2023, this
study aligns with the ongoing push for reproducible and representative
forensic testing, ensuring that performance metrics capture both clas­
sical and emerging manipulation scenarios.
3.5. Summary and research gap
Attention-based architectures have significantly advanced manipu­
lation detection by capturing long-range dependencies and improving
interpretability. However, most existing approaches focus on either bi­
nary authenticity classification or qualitative localisation without
providing a consistent quantitative measure of manipulation severity.
The proposed VISION-ATTENTION ANOMALY SCORING (VAAS) framework
bridges this gap by introducing a dual-module architecture that fuses
global anomaly scoring from ViTs with local tamper segmentation via
SegFormer, providing both quantitative and interpretable forensic
evidence.

3.2. Attention and transformer-based forensic models
The integration of attention mechanisms and transformer architec­
tures has recently reshaped image forensics by enabling models to
reason over spatial and semantic contexts. Attention mechanisms
selectively highlight informative regions, facilitating the identification
of subtle anomalies in lighting, texture, and structure. Vision trans­
formers (ViTs), which employ self-attention to model global relation­
ships between image patches, are particularly suitable for analysing
semantic coherence in tampered imagery (Ma et al., 2023).
Hao et al. (2021) introduced TransForensics, a dense transformer
encoder model that captures multi-scale patch dependencies for forgery
localisation. Wang et al. (2022) proposed ObjectFormer, which applies
attention-guided feature fusion for pixel-level tampering detection. Xia
et al. (2024) further incorporated hierarchical attention to combine
convolutional and transformer features, while Shi et al. (2024) com­
bined dual attention and edge supervision to refine boundary precision.
These studies highlight the effectiveness of attention in learning
semantically rich and interpretable representations.

4. Methodology
The proposed methodology, VISION-ATTENTION ANOMALY SCORING FOR
IMAGE INTEGRITY (VAAS), is formulated as an anomaly detection and
localisation framework rather than a conventional classification or
segmentation task. Instead of relying solely on class labels or pixel-level
masks, the approach focuses on learning the intrinsic spatial consistency
of natural images and identifying deviations introduced by classical or
AI-driven manipulations.
The system integrates two complementary modules, as illustrated in
Fig. 2: (1) a full-image attention-based module (Fx) for global anomaly
detection, and (2) a patch-level self-consistency module (Px) for local
anomaly localisation. Together, these modules quantify spatial irregu­
larities such as texture discontinuities, lighting inconsistencies, and
unnatural object alignments. The outputs from both modules are fused
through a Hybrid Weighted Scoring Mechanism (HSM), which integrates
the global and local anomaly cues into a single, interpretable integrity
score. This formulation allows the system to balance attention-driven
global coherence with patch-level spatial irregularities, providing both
quantitative and visual evidence of tampering for digital forensic
investigators.

3.3. Cross-attention for contextual consistency
Cross-attention mechanisms extend self-attention by allowing
feature interactions across distinct representation spaces, such as global
and local features or authentic and manipulated embeddings. This
design enhances contextual reasoning by enabling the model to compare
and align heterogeneous cues that reflect semantic consistency in
authentic imagery. Chen et al. (2021) introduced CrossViT, a trans­
former model that integrates multi-scale patch tokens through
cross-attention, effectively combining fine-grained and global repre­
sentations. More recently, Munawar and Oussalah (2025) proposed an
explainable dual-stream attention network for image forgery detection
and localisation, employing contrastive learning to strengthen feature
discrimination between genuine and tampered regions. These de­
velopments demonstrate that cross- and dual-attention frameworks can
bridge complementary information across feature domains, motivating
the global–local interaction strategy adopted in the VAAS architecture.

4.1. Full-image consistency module (fx)
The Fx module employs a Vision Transformer (ViT) to extract deep
representations from the entire image and assess its global spatial
coherence. Unlike patch-based methods, this module treats the image
holistically and derives a global anomaly score based on the intensity
and spatial distribution of its attention maps. These maps highlight re­
gions where the model focuses most during inference, providing
explainable cues for identifying structural irregularities.
For reproducibility, the Fx module uses attention maps extracted
from the final four encoder layers of ViT-Base, averaged across all
heads. Images are first resized to 224 × 224 and normalised using
ImageNet statistics. The attention tensor is downsampled to the original
image resolution using bilinear interpolation before scoring. All exper­
iments were run with a fixed random seed to ensure consistency across
training and evaluation.
The process includes:

3.4. Datasets and forensic benchmarking
Benchmark datasets have been essential for the progress of image
forensics (Thomson et al., 2025). The CMFD dataset (Christlein et al.,
2012) provided an early benchmark for copy–move detection under
geometric distortions. CASIA v2.0 by Dong et al. (2013) introduced
splicing and compositing manipulations, enabling the development of
robust segmentation-based forensic networks (Bappy et al., 2019; Zhou
et al., 2018). More recently, the Digital Forensics 2023 Dataset for Image
Forgery Detection (DF2023) by Fischinger and Boyer (2023b) extended
this paradigm by including AI-generated and hybrid manipulations from
diffusion and style-transfer pipelines, thereby supporting large-scale
benchmarking with fine-grained annotations.

1. Feature Extraction: The input image is passed through the ViT,
which generates multi-layer attention maps reflecting contextual
relationships between image patches.
3

O. Bamigbade et al.

Forensic Science International: Digital Investigation 56 (2026) 302063

Fig. 2. Overview of the VAAS framework showing the data flow through its training (left) and inference (right) stages. The architecture integrates global attention
analysis and local consistency estimation to detect spatial inconsistencies indicative of image manipulation.

Fig. 3. Qualitative visualisation of VAAS on CASIA v2.0 (top two rows) and DF2023 (bottom two rows). Columns (1)–(6) show: input image, ground-truth mask,
binary Px output, Px heatmap overlay, Fx attention overlay, and final hybrid anomaly score. High-anomaly samples (top rows) exhibit precise localisation and strong
global cues, while mid-range samples show balanced but diffuse attention, illustrating the complementary interaction between Px and Fx.

4

O. Bamigbade et al.

Forensic Science International: Digital Investigation 56 (2026) 302063

2. Attention Analysis: In authentic images, attention distributions are
typically smooth and semantically consistent. Manipulated regions
introduce irregular attention activations due to inconsistent feature
interactions.
3. Statistical Scoring: The mean (μ) and standard deviation (σ) of
attention activations are compared with reference distributions
computed from authentic samples.

SegFormer, and α ∈ [0, 1] controls the relative influence of global versus
local cues.
A higher α places greater emphasis on global attention deviations,
suitable for manipulations that alter semantic structure or lighting
coherence, while a lower α prioritises local texture and boundary ir­
regularities captured by the Px module. This adaptive formulation al­
lows α to be determined empirically or through cross-validation,
depending on dataset characteristics or the forensic context in which the
model is deployed. Validation analysis or threshold sweeping can
further refine α to ensure that the hybrid metric adapts effectively to
diverse manipulation types.
This weighted fusion provides a continuous, interpretable measure of
image integrity. The resulting score not only reflects the likelihood of
manipulation but also preserves sensitivity to both subtle global
anomalies and fine-grained local inconsistencies, making it particularly
useful for forensic evaluation, where interpretability and balanced evi­
dence integration are critical.
In addition to the weighted formulation, a harmonic variant of the
fusion function was evaluated to explore whether penalising disagree­
ment between the two modules (Fx and Px) enhances detection
robustness. The comparative performance of these two fusion strategies
is presented in Section 5.2. The harmonic variant can be expressed as:

The global anomaly score SF is expressed as:
SF =

|μ − μref |

σref

(1)

where μref and σref represent the reference statistics derived from
authentic attention distributions. A higher SF indicates a stronger devi­
ation from natural spatial coherence, which corresponds to a higher
likelihood of tampering. To evaluate the influence of transformer rep­
resentation depth on anomaly estimation, alternative backbones such as
vit-large-patch16-224, swin-base-patch4-window7-224,
and dinov2-base were also examined. This analysis, detailed in Sec­
tion 5.1, assesses whether global attention granularity impacts the sta­
bility of the forensic anomaly score SF.
4.2. Patch-based self-consistency module (Px)

2
1
+
SF
SP

Sharmonic
= 1
H

The Px module performs local anomaly detection by analysing selfconsistency among non-overlapping image patches. The input image I
is divided into patches Pi of size k × k. Each patch is embedded using a
SegFormer-based encoder to obtain its latent representation F(Pi ).
Local irregularities are quantified by measuring the cosine similarity
between each patch and its spatial neighbours:
F(Pi )⋅F(Pj )
⃦⃦
⃦
Sim(Pi , Pj ) = ⃦
⃦F(Pi )⃦⃦F(Pj )⃦

4.4. Loss functions and training strategy
The Px module is trained in a supervised manner using the available
manipulation masks, while the Fx module provides attention guidance
to enhance spatial coherence. A composite segmentation loss combines
Binary Cross-Entropy (BCE), Dice, and focal components to balance
pixel-level accuracy and region overlap:

(2)

Low similarity indicates contextual deviation, and the anomaly score
per patch is defined as:
1 ∑
SP (i) = 1 −
Sim(Pi , Pj )
N j∈N (i)

L seg = λbce L BCE + λdice L Dice + λfocal L Focal

M
1 ∑
SP (i)
M i=1

(3)

L fx = 1 − cos(FPx , FFx )

(8)

The influence of this alignment is scaled by a regularisation coeffi­
cient ωfx, producing the total loss:
L total = L seg + ωfx L fx

(4)

(9)

where L BCE denotes the binary cross-entropy loss, L Dice is the Dice
overlap loss, and L Focal represents the focal loss used to handle class
imbalance. λbce, λdice, and λfocal are scalar weighting factors controlling
each term's contribution within the segmentation loss L seg . The ωfx is a
regularisation coefficient that balances the influence of the attention
alignment term relative to the segmentation objective, yielding the total
optimisation loss L total .
A moderate value of ωfx = 0.1 yielded an optimal balance between
feature stability and mask sharpness across datasets. The Fx module it­
self remains frozen during training, serving only as a provider of pretrained attention maps that guide Px feature refinement through crossattention mechanism.
VAAS performance was assessed using a number of both detection
and localisation metrics. The selected ones are (1) F1-Score: Balances
precision and recall for manipulation detection, (2) IoU (Intersectionover-Union): Evaluates the overlap between predicted and ground-truth
masks, and (3) Precision and Recall: Capturing trade-offs between false
alarms and missed detections.

where M is the number of patches. This representation encodes how well
local textures and structures align with their spatial context.
In practice, the SegFormer encoder outputs 256-dimensional patch
embeddings. Image patches were extracted with a fixed patch size of 32
× 32 and non-overlapping stride. All ground-truth masks were resized to
224 × 224 using nearest-neighbour interpolation to preserve boundary
sharpness. Patch-level anomaly scores were upsampled back to full
resolution before fusion with the global score.
4.3. Hybrid anomaly scoring mechanism (HSM)
The final anomaly score integrates both global and local evidence of
tampering obtained from the Fx and Px modules, respectively. Rather
than relying on a single cue, the proposed framework employs a
weighted fusion strategy that balances global contextual irregularities
with local spatial inconsistencies. The unified score, denoted as SH, is
defined as:
SH = α SF + (1 − α) SP

(7)

To further align the local features of Px with the global context learnt
by Fx, an attention alignment term is introduced:

where N (i) is the neighbourhood of patch i, and N is its size. The
final patch-level anomaly score is aggregated as:
SP =

(6)

4.5. Dataset overview and implementation

(5)

The VAAS framework was evaluated on two established forensic
datasets: CASIA v2.0 and DF2023, which were chosen to represent both

where SF is the global attention-based anomaly score from the Vision
Transformer, SP is the local patch-level consistency score from the
5

O. Bamigbade et al.

Forensic Science International: Digital Investigation 56 (2026) 302063

traditional and AI-driven manipulation scenarios.
CASIA v2.0 (Dong et al., 2013) is a foundational benchmark for
evaluating splicing and compositing detection methods. It contains
approximately 12,614 images, including 7491 authentic and 5123
tampered samples, each with a corresponding ground-truth mask. The
dataset spans diverse real-world scenes and manipulation styles,
including post-processing variations. All images were resized to 224 ×
224 and normalised to [0,1]. The Px module was trained using
tampering masks, while the Fx module derived reference attention sta­
tistics (μref, σ ref) from authentic samples.
DF2023 (Fischinger and Boyer, 2023b) was designed to benchmark
forensic methods against AI-generated and hybrid manipulations. The
full dataset contains approximately one million forged images distrib­
uted across four manipulation types: 100K removal, 200K enhancement,
300K copy–move, and 400K splicing operations. The first work on
DF2023 was conducted by Fischinger and Boyer (2023a), who were only
able to use a fraction of the dataset due to its computational intensity,
and their reported results focused on other related manipulation data­
sets rather than DF2023 itself. In this study, a stratified 10 % subset
(approximately 100k images) of DF2023 was used to make training
computationally feasible. The subset was sampled proportionally from
each manipulation category (100k removal, 200k enhancement, 300k
copy–move, 400k splicing), preserving the original distribution. All
experiments used an 80/20 train–validation split, and the exact file
indices used for sampling are included in the public code repository for
full reproducibility. The validation set serves as the testing set for all
reported metrics, and no DF2023 samples outside this split were used
during evaluation.
The dataset's diversity and generative content make it particularly
suitable for evaluating VAAS's hybrid anomaly scoring behaviour under
semantically coherent but spatially inconsistent manipulations.
Experiments were conducted in PyTorch using the Hugging Face
Transformers library. The Fx module employed google/vit-basepatch16-224-in21k, while Px used nvidia/segformer-b1finetuned-ade-512-512. Both models were trained using the Adam
optimiser (learning rate 1 × 10− 4, and a batch size of 8). The composite
loss combines BCE, Dice, and focal components, balanced by coefficients
λdice = 0.7 and λfocal = 1.0, as detailed in Section 4. Training was per­
formed on an NVIDIA GeForce RTX 5090 GPU (34 GB VRAM) for 80
epochs, with checkpoints selected based on validated F1-scores. Evalu­
ation metrics included F1-score, IoU, precision, and recall.
Although Hotels-50K (Stylianou et al., 2019) is not an image
manipulation dataset, it remains operationally relevant in forensic
contexts involving large-scale image verification. Recent work using this
dataset on colour-sensitive embedding design for indoor scene recog­
nition (Bamigbade et al., 2025) and indoor multimedia geolocation
(Aftab et al., 2026) highlights the broader interest in developing inter­
pretable visual representations for digital forensic investigative tasks.
While this prior work is not directly related to tampering detection, it
motivates future exploration of how anomaly scoring methods, such as
VAAS, might support integrity assessment within similar real-world
pipelines.
In summary, the methodological design of VAAS provides a balanced
framework for evaluating both interpretability and forensic robustness
across traditional and AI-generated manipulations. With the hybrid
anomaly scoring mechanism and Fx-guided training in place, the
following section presents the empirical evaluation of VAAS. Results are
reported for both datasets, highlighting quantitative performance (F1,
IoU) and qualitative interpretability through attention-guided visual­
isations. These analyses demonstrate how the proposed architecture
generalises across manipulation types and scales while maintaining
explainable decision pathways critical for digital forensic reliability.

justify the configuration adopted in the main experiments, a series of
component analysis studies were conducted. Each experiment isolates a
specific architectural or hyperparameter component and evaluates its
impact on detection accuracy, localisation precision, and interpret­
ability. The following four analyses collectively provide empirical
insight into the contributions and interactions of the major components:
1. Fx Backbone Variation: evaluates the effect of transformer depth
and architecture type on global anomaly estimation.
2. Fusion Mechanism (Weighted vs Harmonic): examines alterna­
tive formulations of the hybrid scoring function and their stability
across datasets.
3. Effect of Fx-Guided Regularisation: investigates how the strength
of attention-based guidance during training influences Px feature
alignment and segmentation fidelity.
4. Dataset-Specific Sensitivity of α: analyses how the weighting factor
controlling Fx–Px contributions adapts across manipulation
domains.
Together, these studies clarify the internal dynamics of VAAS,
highlighting how attention coupling and fusion weighting jointly
contribute to balanced performance and interpretability in forensic
image analysis.
5.1. Fx Backbone Variation
To examine the influence of transformer capacity and attention
granularity on global anomaly estimation, the default ViT-Base
backbone was compared with ViT-Large, Swin-Base, and DINOv2Base. Models with deeper or hierarchical attention (e.g., Swin) pro­
duced smoother anomaly heatmaps but exhibited marginal F1
improvement (+1.3 % on DF2023). The standard ViT-Base provides the
best trade-off between interpretability and computational efficiency,
confirming its suitability for Fx within the VAAS framework. Fig. 4 vi­
sualises both the quantitative and qualitative impact of different Fx
backbones, illustrating how transformer depth and hierarchy affect
anomaly sharpness and interpretability.
5.2. Fusion Mechanism: weighted vs harmonic
A comparison was conducted between the linear weighted fusion
(Eq. (5) and the harmonic variant (Eq. (6) used to combine global and
local anomaly scores. Both the weighted and harmonic formulations
were examined by sweeping the weighting factor α in the range [0.3,
0.8]. As shown in Fig. 5, performance improves steadily with increasing
α up to approximately 0.6, where the influence of global attention cues
(Fx) and local patch consistency (Px) is balanced. Beyond this point,
excessive reliance on the global term leads to marginal performance
saturation. The harmonic variant demonstrates smoother, more stable
behaviour across datasets, indicating its robustness when Fx and Px
provide conflicting evidence. This confirms that harmonic fusion not
only stabilises decision confidence but also enhances interpretability by

5. VAAS component analysis

Fig. 4. Fx backbone component analysis showing F1-scores on DF2023 and
CASIA v2.0. ViT-Base offers the best balance between accuracy and efficiency.

To better understand the design choices underlying VAAS and to
6

Forensic Science International: Digital Investigation 56 (2026) 302063

O. Bamigbade et al.

transparent deployment.
6. Results
The proposed VAAS framework was evaluated on two complemen­
tary forensic benchmarks: the DF2023 dataset, which represents largescale and heterogeneous image manipulations, and the CASIA v2.0
Image Tampering Detection Evaluation Database, which represents
classical splicing and compositing forgeries. To ensure contextual
benchmarking, representative state-of-the-art methods from both
transformer-based and convolutional paradigms were selected, covering
the spectrum of modern forensic detection strategies.
For CASIA v2.0, four published approaches were used for compari­
son: the attention-based Dual-Stream Attention Network (DSCL-Net) by
Munawar and Oussalah (2025), the visually guided VASLNet (Yadav and
Vishwakarma, 2024), the hybrid convolutional–transformer model
Hybrid CNN–Transformer (Sharma et al., 2025), and the hierarchical
fine-grained localisation framework HiFi-IFDL (Guo et al., 2023). These
baselines collectively represent a balanced mix of classical and deep
attention-based strategies for manipulation detection and localisation.
For DF2023, DF-Net (Fischinger and Boyer, 2023a) served as the
dataset's reference model. Although the dataset was introduced in
Fischinger and Boyer (2023b), no prior work has reported comprehen­
sive quantitative metrics directly on DF2023. The DF-Net study evalu­
ated cross-dataset performance on other benchmarks only. In this study,
VAAS establishes the first reproducible detection and localisation met­
rics on DF2023, trained and validated on 10 % (Over 100k) of the full
corpus with an 80/20 split, preserving the manipulation-type proportions
described in the official release.
Table 1 presents the comparative quantitative results. On CASIA
v2.0, VAAS achieved performance comparable to or exceeding attentionand CNN-based models, particularly in localisation (IoU), confirming
the robustness and generalisability of its hybrid scoring mechanism
across manipulation types. On DF2023, VAAS provides the first bench­
marked reference for both detection and localisation, supporting future
evaluation of generative and hybrid tampering detection methods.
Fig. 3 illustrates qualitative comparisons between manipulated im­
ages, ground-truth masks, and VAAS-generated attention maps. The Fx
module (ViT) captures global inconsistencies in illumination and scene
coherence, while the Px module (SegFormer) isolates fine-grained ar­
tefacts such as blending edges or inpainting seams. Their weighted
fusion produces anomaly maps that align closely with manipulated re­
gions, reinforcing both interpretability and forensic reliability.

Fig. 5. Anomaly scoring fusion component analysis showing F1-scores across α.
Harmonic fusion varies smoothly and remains stable across datasets, while
weighted fusion peaks near α = 0.6.

weighting the agreement between global and local anomaly cues.
5.3. Effect of Fx-guided regularisation
To assess the influence of the Fx-guided attention alignment term,
the regularisation coefficient ωfx was varied in {0, 0.05, 0.1, 0.2}(see
Fig. 6). A moderate value (ωfx = λfx = 0.1) achieved an optimal balance
between local boundary precision and global attention coherence.
Setting ωfx = 0 led to noisier segmentation masks, indicating that Fx cues
help stabilise Px feature learning, while higher weights (ωfx > 0.2)
caused oversmoothing and reduced edge fidelity.
5.4. Dataset-Specific Sensitivity of α
The weighting factor α controlling the contribution of Fx and Px
scores was evaluated to identify dataset-specific trends. For CASIA v2.0,
optimal α values lie between 0.4 and 0.6, reflecting a balanced reliance
on local and global cues. For DF2023, higher values (α ≈ 0.7) yielded
superior results due to the predominance of semantically coherent yet
globally inconsistent manipulations produced by generative models.
This analysis highlights the adaptability of VAAS across domains, with α
serving as an interpretable control parameter for forensic sensitivity.
In summary, the component-wise analysis consistently supports the
design rationale of VAAS. The ViT-Base backbone offers a strong balance
between representational capacity and interpretability for global
anomaly estimation, while the weighted fusion strategy provides stable
performance across datasets. Fx-guided regularisation improves feature
alignment and reduces noise in Px outputs, highlighting the value of
cross-attention coupling during training. Finally, the dataset-dependent
behaviour of α shows that anomaly sensitivity can be tuned to different
manipulation characteristics, supporting both adaptability and

6.1. Quantitative evaluation
As summarised in Table 1, VAAS achieved an F1-score of 94.9 % and
Table 1
Quantitative comparison of VAAS with representative state-of-the-art methods
on DF2023 and CASIA v2.0. Baseline values are taken directly from the cited
publications. “–” indicates metrics not reported. VAAS achieves comparable or
superior performance in both detection (F1) and localisation (IoU) metrics.
Dataset

Method

Precision
(%)

Recall
(%)

F1
(%)

IoU
(%)

CASIA
v2.0

DSCL-Net (Dual-Stream
Attn.) (Munawar and
Oussalah, 2025)
VASLNet (Yadav and
Vishwakarma, 2024)
Hybrid CNN–Transformer (
Sharma et al., 2025)
HiFi-IFDL (Guo et al., 2023)
VAAS

–

–

92.7

–

–

–

91.9

85.1

–

–

84.0

82.0

99.5
93.5

–
94.8

97.4
94.1

61.6
89.0

DF-Net (Fischinger and
Boyer, 2023a)
VAAS

–

–

–

–

95.9

94.2

94.9

91.1

DF2023

Fig. 6. Effect of Fx-guided regularisation weight λfx. Moderate values (λfx =
0.1) achieve the best trade-off between boundary precision and atten­
tion coherence.
7

O. Bamigbade et al.

Forensic Science International: Digital Investigation 56 (2026) 302063

an IoU of 91.1 % on DF2023, outperforming the dataset baseline. This
improvement reflects the effectiveness of integrating global attention
cues (SF) with patch-level self-consistency (SP), allowing the framework
to capture both semantic and structural inconsistencies typical of
generative manipulations. On CASIA v2.0, VAAS attained an F1-score of
94.1 % and an IoU of 89.0 %, exceeding most attention-based and hybrid
CNN–Transformer methods. Although HiFi-IFDL reported a slightly
higher F1 due to its hierarchical refinement, its lower IoU suggests less
stable localisation. In contrast, VAAS maintains a stronger balance be­
tween detection accuracy and boundary precision, confirming the
advantage of its hybrid scoring design.
To examine how the hybrid weighting parameter α influences per­
formance, a threshold sweep was performed across multiple global–local
ratios (Fig. 7). Both F1 and IoU peaked around α = 0.6, indicating that a
moderate emphasis on the global attention term (SF) provides the most
consistent trade-off between detection and localisation accuracy.

trends reported in Section 6. Higher anomaly scores correspond to
confident, spatially consistent manipulations, while moderate scores
reflect partial or ambiguous edits. This coherence between anomaly
intensity and visual interpretability reinforces VAAS as a transparent
and reliable forensic tool.
6.3. Interpretability and stability
Beyond quantitative accuracy, VAAS emphasises transparent and
stable forensic reasoning. The Fx attention heads generate semantically
meaningful anomaly maps that align with manipulated and disrupted
semantic regions, enabling human analysts to visually verify model in­
ferences. For authentic images, attention patterns remain compact and
coherent, whereas manipulated samples exhibit fragmented or diffuse
activation, clearly indicating the presence of anomalies. This interpret­
ability supports explainable decision-making, an essential requirement
for forensic and legal reliability.
The framework also demonstrates adaptive stability across manipu­
lation scales. In CASIA v2.0, where tampering is typically localised, Px
dominates the anomaly response; in DF2023, which contains global se­
mantic modifications, Fx contributes more strongly. The adaptive
weighting parameter α dynamically balances these contributions,
allowing VAAS to generalise effectively across datasets and manipula­
tion types.

6.2. Qualitative analysis
Fig. 3 illustrates the qualitative behaviour of the proposed VAAS
framework across both CASIA v2.0 and DF2023. Each figure is organised
into four rows and six columns: the top two rows correspond to CASIA
samples, and the bottom two rows correspond to DF2023. Columns (1)–
(6) show the input image, ground-truth mask, binary Px output, Px
heatmap overlay, Fx attention overlay, and the hybrid anomaly score
visualisation.
Two representative anomaly levels are presented for each dataset:
high and mid-range scores. In high-anomaly cases, Px generates seg­
mentation masks that closely match the ground truth, while Fx high­
lights broader contextual irregularities, such as illumination shifts or
semantic imbalances. This behaviour aligns with the design intent of the
hybrid scoring, where the global stream (SF) augments the precision of
the patch-based stream (SP) to produce spatially coherent forensic
evidence.
For mid-range anomalies, Px still localises the main tampered re­
gions; however, minor deviations occur at edges or texture transitions.
Meanwhile, Fx continues to detect subtle global inconsistencies, occa­
sionally extending attention to adjacent or contextually related areas; an
effect consistent with global semantic reasoning. Together, these re­
sponses illustrate how VAAS integrates complementary cues: Px pro­
vides spatial fidelity, while Fx contributes contextual awareness.
It is important to note that the Fx attention map is not intended to
replicate the pixel-level manipulation mask. Vision Transformer atten­
tion highlights regions whose semantic or structural coherence is dis­
rupted by a manipulation, which may extend beyond the exact tampered
boundary. This behaviour is consistent with global reasoning: the model
focuses on objects or regions that become contextually inconsistent once
an edit is introduced. These global cues complement the precise spatial
localisation produced by Px, and together they form a more complete
explanation of anomaly evidence.
Overall, the qualitative patterns in Fig. 3 mirror the quantitative

6.4. Summary of findings
In summary, VAAS delivers competitive or superior performance
across both benchmarks while maintaining interpretability and stability.
Its hybrid scoring mechanism balances the fine-grained sensitivity of
patch-level localisation with the contextual depth of transformer-based
attention. By unifying these complementary cues, VAAS provides
explainable, reproducible, and scalable forensic reasoning; key traits for
operational digital evidence verification and domain-wide deployment.
7. Discussion
Building on the findings in Section 6, this section interprets the
performance and architectural behaviour of the proposed VAAS frame­
work. The discussion focuses on how the interaction between the fullimage attention module (Fx), the patch-level segmentation module
(Px), and the hybrid scoring mechanism (HSM) jointly enhances detec­
tion accuracy and interpretability. It also outlines the implications of
these results for digital forensic practice, where explainable and repro­
ducible evidence generation is a critical requirement.
As shown in Section 6, the quantitative results indicate that a mod­
erate global–local weighting (α ≈ 0.6) achieves an optimal balance be­
tween detection accuracy and localisation precision. This finding aligns
with the qualitative observations from inference, where the Fx and Px
modules demonstrate complementary behaviour in highlighting se­
mantic and structural inconsistencies.
The experiments confirm that VAAS attains strong quantitative
performance while preserving interpretability in localising manipulated
regions. The hybrid scoring mechanism effectively integrates global
anomaly cues from Fx with spatially detailed predictions from Px,
enabling the detection of subtle or spatially inconsistent manipulations
that are often overlooked by conventional methods. The resulting
anomaly maps provide visual explanations that clarify the model's de­
cision boundaries, which are particularly valuable in forensic analysis
requiring traceable reasoning.
The transformer-based Fx backbone contributes to the stability and
expressiveness of global attention representations. By offering a
consistent contextual signal to guide segmentation, Fx ensures that
anomaly localisation from Px aligns with actual tampered regions. In
turn, the Px module, implemented with SegFormer, refines the local­
isation through patch-level supervision. Together, these modules form a

Fig. 7. Threshold sweep analysis of the hybrid scoring weight α in SH = αSF +
(1 − α)SP with a performance peaks around α = 0.6.
8

Forensic Science International: Digital Investigation 56 (2026) 302063

O. Bamigbade et al.

cross-scale interaction that balances detection sensitivity, localisation
precision, and interpretability; three essential pillars for operational
forensic reliability.
Overall, these findings demonstrate that VAAS not only advances
quantitative performance but also provides a principled foundation for
interpretable and scalable forensic AI systems, motivating further
exploration of its limitations and future extensions.

Christlein, V., Riess, C., Jordan, J., Riess, C., Angelopoulou, E., 2012. An evaluation of
popular copy-move forgery detection approaches. IEEE Trans. Inf. Forensics Secur. 7
(6), 1841–1854.
Dong, J., Wang, W., Tan, T., 2013. CASIA image tampering detection evaluation
database. In: 2013 IEEE China Summit and International Conference on Signal and
Information Processing, pp. 422–426.
Du, X., Hargreaves, C., Sheppard, J., Anda, F., Sayakkara, A., Le-Khac, N.-A., Scanlon, M.,
2020. SoK: exploring the state of the art and the future potential of artificial
intelligence in digital forensic investigation. In: Proceedings of the 15th
International Conference on Availability, Reliability and Security. ARES ’20.
Association for Computing Machinery, New York, NY, USA. https://doi.org/
10.1145/3407023.3407068.
Feng, D., Lu, X., Lin, X., 2020. Deep detection for face manipulation. In: Yang, H.,
Pasupa, K., Leung, A.C.-S., Kwok, J.T., Chan, J.H., King, I. (Eds.), Neural Information
Processing. Springer International Publishing, Cham, pp. 316–323.
Fischinger, D., Boyer, M., 2023a. DF-Net: the digital forensics network for image forgery
detection. In: Corcoran, P. (Ed.), Proceedings of 25th Irish Machine Vision and Image
Processing Conference (IMVIP 2023), pp. 120–127. URL. https://iprcs.github.io
/pdf/IMVIP2023_Proceeding.pdf.
Fischinger, D., Boyer, M., 2023b. DF2023: the digital forensics 2023 dataset for image
forgery detection. In: Corcoran, P. (Ed.), Proceedings of 25th Irish Machine Vision
and Image Processing Conference (IMVIP 2023), pp. 128–135. URL. https://iprcs.
github.io/pdf/IMVIP2023_Proceeding.pdf.
Guarnera, L., Giudice, O., Battiato, S., 2024. Mastering deepfake detection: a cuttingedge approach to distinguish GAN and diffusion-model images. ACM Trans.
Multimed Comput. Commun. Appl 20 (11). https://doi.org/10.1145/3652027.
Guo, X., Liu, X., Ren, Z., Grosz, S., Masi, I., Liu, X., 2023. Hierarchical fine-grained image
forgery detection and localization. In: 2023 IEEE/CVF Conference on Computer
Vision and Pattern Recognition (CVPR), pp. 3155–3165.
Hao, J., Zhang, Z., Yang, S., Xie, D., Pu, S., 2021. TransForensics: image forgery
localization with dense self-attention. In: 2021 IEEE/CVF International Conference
on Computer Vision (ICCV), pp. 15035–15044.
Hargreaves, C., Breitinger, F., Dowthwaite, L., Webb, H., Scanlon, M., 2024. DFPulse: the
2024 digital forensic practitioner survey. Forensic Sci. Int.: Digit. Invest. 51, 301844.
URL. https://www.sciencedirect.com/science/article/pii/S2666281724001719.
Kingra, S., Aggarwal, N., Kaur, N., 2024. SFormer: an end-to-end spatio-temporal
transformer architecture for deepfake detection. Forensic Sci. Int.: Digit. Invest. 51,
301817. URL. https://www.sciencedirect.com/science/article/pii/S2666281
724001410.
Li, Y., Chang, M.-C., Lyu, S., 2018. Ictu oculi: exposing AI created fake videos by
detecting eye blinking. In: 2018 IEEE International Workshop on Information
Forensics and Security (WIFS), pp. 1–7.
Li, Y., Yang, X., Sun, P., Qi, H., Lyu, S., 2020. Celeb-DF: a large-scale challenging dataset
for DeepFake forensics. In: 2020 IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), pp. 3204–3213.
Ma, X., Du, B., Jiang, Z., Hammadi, A.Y.A., Zhou, J., 2023. IML-ViT: benchmarking
image manipulation localization by vision transformer. arXiv preprint arXiv:
2307.14863. URL. https://arxiv.org/abs/2307.14863.
Mahdian, B., Saic, S., 2008. Detection of resampling supplemented with noise
inconsistencies analysis for image forensics. In: 2008 International Conference on
Computational Sciences and its Applications, pp. 546–556.
Matern, F., Riess, C., Stamminger, M., 2019. Exploiting visual artifacts to expose
deepfakes and face manipulations. In: 2019 IEEE Winter Applications of Computer
Vision Workshops (WACVW), pp. 83–92.
Monika, Bansal, D., Passi, A., 2021. Image forensic investigation using discrete cosine
transform-based approach. Wirel. Pers. Commun. 119 (4), 3241–3253. https://doi.
org/10.1007/s11277-021-08396-1.
Munawar, M., Oussalah, M., 2025. Explainable dual-stream attention network for image
forgery detection and localisation using contrastive learning. IET Radar, Sonar
Navig. 19 (1), e70064.
Nguyen, H.H., Yamagishi, J., Echizen, I., 2022. Capsule-Forensics Networks for Deepfake
Detection. Springer International Publishing, Cham, pp. 275–301. https://doi.org/
10.1007/978-3-030-87664-7_13.
Pandey, R.C., Singh, S.K., Shukla, K.K., Agrawal, R., 2014. Fast and robust passive copymove forgery detection using SURF and SIFT image features. In: 2014 9th
International Conference on Industrial and Information Systems (ICIIS), pp. 1–6.
Sharma, S., Singh, B.K., Garg, H., 2025. Robust image forgery localization using hybrid
CNN-transformer synergy based framework. Comput. Mater. Continua (CMC) 82 (3),
4691–4708.
Shi, C., Wang, C., Zhou, X., Qin, Z., 2024. DAE-Net: dual attention mechanism and edge
supervision network for image manipulation detection and localization. IEEE Trans.
Instrum. Meas. 73, 1–17.
Singh, J., 2022. Deepfakes: the threat to data authenticity and public trust in the age of
AI-Driven manipulation of visual and audio content. J. AI-Assisted Scientific Discov.
2 (1), 428–467.
Spichiger, H., Adelstein, F., 2025. Preserving meaning of evidence from evolving
systems. Forensic Sci. Int.: Digit. Invest. 52, 301867. DFRWS EU 2025 - Selected
Papers from the 12th Annual Digital Forensics Research Conference Europe. URL.
https://www.sciencedirect.com/science/article/pii/S266628172500006X.
Stylianou, A., Xuan, H., Shende, M., Brandt, J., Souvenir, R., Pless, R., 2019. Hotels-50K:
a global hotel recognition dataset. In: Proceedings of the Thirty-Third AAAI
Conference on Artificial Intelligence and Thirty-First Innovative Applications of
Artificial Intelligence Conference and Ninth AAAI Symposium on Educational
Advances in Artificial Intelligence. AAAI’19/IAAI’19/EAAI’19. AAAI Press. https://
doi.org/10.1609/aaai.v33i01.3301726.

7.1. Limitations and future work
While VAAS demonstrates strong interpretability and competitive
accuracy, several limitations remain. First, part of the evaluation was
conducted on subsets of DF2023, which may not fully capture the di­
versity and complexity of the manipulations encountered in operational
forensic imagery. A further limitation is that the current evaluation does
not include cross-dataset testing, which is increasingly recognised as a
necessary measure for assessing the robustness of AI-based forensic
systems to unseen distributions.
Although Hotels-50K is not an image manipulation dataset and is not
used in the experiments of this paper, it remains operationally relevant
to digital forensics due to its role in real-world hotel room identification.
Its inclusion is solely to motivate future extensions of VAAS beyond
controlled manipulation benchmarks. In such domains, the goal shifts
from detecting tampering to assessing broader visual integrity and
provenance.
Future work will expand the framework to a wider range of manip­
ulation categories, including composite AI–human edits, and will eval­
uate cross-domain generalisation on unseen datasets. Extending VAAS to
video and multimodal forensics also represents a promising direction.
Methodologically, incorporating adaptive thresholding within the har­
monic scoring mechanism may enhance robustness across varied con­
tent domains. Beyond detection metrics, future research should quantify
interpretability itself by assessing how effectively the generated anom­
aly maps assist human analysts in verifying manipulation evidence.
8. Conclusion
This paper presents VAAS, a vision–attention anomaly scoring
framework for detecting and localising image manipulations in digital
forensics. The approach combines a global attention analysis (Fx) with a
local segmentation process (Px) through a hybrid harmonic scoring
mechanism that quantifies image integrity. Evaluations on CASIA v2.0
and DF2023 demonstrate that VAAS achieves strong detection accuracy
and stable localisation while producing interpretable attention heat­
maps that support forensic decision-making.
The findings confirm that transformer-based attention representa­
tions effectively model spatial inconsistencies and that the proposed
hybrid fusion yields a balanced and explainable measure of authenticity.
By embedding interpretability into the detection process, VAAS ad­
vances forensic readiness and supports transparent verification of visual
evidence in investigative and legal contexts. Future extensions will
explore adapting VAAS to high-variability, real-world datasets such as
Hotels-50K, where the focus shifts from manipulation detection to gen­
eral image integrity and provenance verification.
References
Aftab, K., Adams, G., Scanlon, M., 2026. Plug to place: indoor multimedia geolocation
from electrical sockets for digital investigation. Forensic Sci. Int.: Digit. Invest. 56S.
Bamigbade, O., Scanlon, M., Sheppard, J., 2025. Improving image embeddings with
colour features in indoor scene geolocation. IEEE Access 13, 79860–79870.
Bappy, J.H., Simons, C., Nataraj, L., Manjunath, B.S., Roy-Chowdhury, A.K., 2019.
Hybrid LSTM and encoder–decoder architecture for detection of image forgeries.
IEEE Trans. Image Process. 28 (7), 3286–3300.
Chen, C.-F.R., Fan, Q., Panda, R., 2021. CrossViT: cross-attention multi-scale vision
transformer for image classification. In: Proceedings of the IEEE/CVF International
Conference on Computer Vision (ICCV), pp. 357–366.

9

O. Bamigbade et al.

Forensic Science International: Digital Investigation 56 (2026) 302063

Thomson, M., McKeown, S., Macfarlane, R., Leimich, P., 2025. Exploring dataset
diversity for GenAI image inpainting localisation in digital forensics. In: Proceedings
of the Digital Forensics Doctoral Symposium. DFDS ’25. Association for Computing
Machinery, New York, NY, USA. https://doi.org/10.1145/3712716.3712724.
Verdoliva, L., 2020. Media forensics and DeepFakes: an overview. IEEE J. Selected Topics
Signal Processing 14 (5), 910–932.
Wang, J., Wu, Z., Chen, J., Han, X., Shrivastava, A., Lim, S.-N., Jiang, Y.-G., 2022.
ObjectFormer for image manipulation detection and localization. In: 2022 IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2354–2363.
Wu, Y., AbdAlmageed, W., Natarajan, P., 2019. ManTra-Net: manipulation tracing
network for detection and localization of image forgeries with anomalous features.
In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),
pp. 9535–9544.

Xia, X., Su, L.C., Wang, S.P., Li, X.Y., 2024. DMFF-Net: double-stream multilevel feature
fusion network for image forgery localization. Eng. Appl. Artif. Intell. 127, 107200.
URL. https://www.sciencedirect.com/science/article/pii/S0952197623013842.
Yadav, A., Vishwakarma, D.K., 2024. A Visually Attentive Splice Localization Network
with Multi-Domain Feature Extractor and Multi-Receptive Field Upsampler arXiv
preprint arXiv:2401.06995Preprint.
Zhou, P., Han, X., Morariu, V.I., Davis, L.S., 2018. Learning rich features for image
manipulation detection. In: 2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pp. 1053–1061.
Zhuang, P., Li, H., Tan, S., Li, B., Huang, J., 2021. Image tampering localization using a
dense fully convolutional network. IEEE Trans. Inf. Forensics Secur. 16, 2986–2999.

10