TL;DR
A detection framework based on noise-invariant contrastive learning to enhance adversarial robustness by aligning feature representations of original and noise-augmented samples, which significantly outperforms conventional adversarial training.
FDT-PC (Frequency Domain Transformation with Perceptual Constraints), a novel method that enhances adversarial transferability across different model architectures, is proposed, which achieves superior black-box attack performance on both CNNs and Vision Transformers, outperforming existing state-of-the-art input transformation methods.
Deepfake technology employs state-of-the-art deep learning to create hyper-realistic synthetic media, which has critical implications for digital authenticity and information security. The spread of such faked content has dire social consequences, facilitating misinformation campaigns, identity theft, and loss of trust in digital media. Today's detection systems experience limitations, such as single-model architectures tending to be unable to generalize the variety of artifact patterns in varying deepfake generation techniques, and most methods are not robust to adversarial attacks. To address these challenges, this study introduces an ensemble-based approach, namely the X-Iv2 Ensemble approach, merging Inception ResNet v2 and Xception Net based on their complementary architectures to enhance feature extraction and classification. The approach utilizes a region-based preprocessing technique, dividing input faces into Upperface, Lowerface, and Eyes regions using facial landmarks on four varied datasets such as DFDC, Celeb-DF, FF++ and SDFVD to ensure robust generalization. Notably, the inclusion of varied races, ages, and genders in the dataset ensures equitable performance across demographics. Apart from these improvements, most deep learning detection frameworks are still black boxes with limited interpretability and practical applications. This approach integrates Explainable AI (XAI) using Integrated Gradients to interpret decision-making processes, while adversarial testing through input perturbations comprehensively assesses the robustness of the model. Extensive assessment metrics such as accuracy (97%), sensitivity (97%), and specificity (96%) illustrate the system's better performance in balancing deepfake detection effectiveness and generalization across datasets.
This paper introduces a new anagram based attack into the literature and trains the model to learn these representations by adding a novel label to the dataset which is used as a feature for training the model.
Deepfake detection remains a challenging problem due to the ever-evolving techniques used to synthesize realistic fake images and videos. Traditional detection methods primarily operate in the pixel domain, often struggling to generalize across different forgery types. In this work, we explore an approach that integrates wavelet-based frequency analysis with deep learning to enhance deepfake detection. By decomposing input images using the Discrete Wavelet Transform (DWT), we extract multi-resolution features that capture both global structures and fine-grained manipulation artifacts; such features are then input to a neural network. Our results show that level-1 DWT features mainly induce a sensitivity-specificity tradeoff: they improve fake-class sensitivity and F1 in cross-forgery settings, while RGB models retain stronger specificity and in-forgery performance. This suggests that wavelet sub-bands expose manipulation cues that are useful for detecting unseen fake classes, but they should not be interpreted as a uniform robustness improvement.
Deepfakes pose growing risks to information integrity, yet many detectors perform well only on the datasets they were trained on and remain opaque to human analysts. A robust, explainable detection framework is presented that combines a CNN backbone for extracting spatial artifacts with an LSTM module for modeling temporal inconsistencies across frames. To make decisions auditable, the architecture incorporates Grad-CAM for spatial heatmaps, SHAP for quantitative feature attribution, and LIME for local surrogate explanations. The system was trained primarily on FaceForensics++ with stratified sampling and augmentation to reduce dataset bias and evaluated on multiple external benchmarks to assess cross-domain generalization. Experimental results show strong detection metrics, such as accuracy of 96.3%, precision of 95.8%, recall of 96.7%, and an F1-score of 96.2%, along with robust performance under JPEG compression, Gaussian noise, and FGSM adversarial attacks. By coupling high detection accuracy with transparent explanations, the proposed approach enhances forensic decision support and increases practical readiness for content verification systems.