FORENSIC ANALYSIS AND DETECTION OF MANIPULATED IMAGES USING DEEP LEARNING TECHNIQUES

Abstract

The detection of image manipulation chains in JPEG-compressed environments re mains among the most consequential and technically demanding problems in the field of digital image forensics. Real-world forgeries rarely involve isolated, single-step processing operations; instead, they are constructed through ordered sequences of ma nipulations, spatial filtering, blurring, resampling, and noise injection which are sub sequently subjected to lossy JPEG compression prior to distribution. This compound character of practical forgeries fundamentally undermines single-operation forensic detectors, which perform well in controlled settings but fail to generalise when succes sive operations systematically obscure earlier forensic evidence. This thesis presents AFNet (AttentionForensicNet), a novel attention-driven dual stream convolutional network that addresses the combined challenges of manipula tion chain detection and lossy compression through strategic integration of the Convo lutional Block Attention Module (CBAM) within a dual-stream forensic framework. AFNet processes 256×256 grayscale image patches through two dedicated streams: the Noise Residual Attention Stream (NRAS), which extracts spatial-domain forensic fingerprints through residual learning, and the Compression Artifact Stream (CAS), which operates in the DCT frequency domain to model JPEG-specific artefacts and their interaction with prior manipulation traces. CBAM is integrated at three comple mentary locations, within every residual block of both streams, after the initial con volutional layers of each stream, and within the cross-stream fusion layer, enabling the network to perform dynamic, content-aware feature selection at each stage of the processing hierarchy. Comprehensive experimental evaluation on the RAISE ALL benchmark demon strates that AFNet achieves 97.17% overall test accuracy, surpassing all evaluated baseline systems including constrained CNNs, multi-scale residual networks, and prior dual-stream architectures. Per-class analysis reveals especially pronounced improve ments on the hardest compound smoothing chains: 97.79% on MF→GB chains and 92.15% on GB→MF chains, compared to 84.10% achieved by the strongest single stream competitor. Cross-dataset evaluation on the DRESDEN and UCID bench marks confirms that the attention-learned features generalise across diverse camera systems and acquisition conditions. Ablation studies validate that every CBAM inte gration point contributes uniquely, and that joint channel-spatial attention is superior to channel-only or spatial-only alternatives. The complete attention augmentation intro duces only 0.08M additional parameters and approximately 2.2% additional inference latency, establishing AFNet as both highly effective and computationally practical.