Deepfakes and impersonation

This paper presents DeepForgeSeal, a deep learning framework for proactive deepfake detection using semi-fragile watermarks. A learnable watermark embedder operates in latent space, and Adversarial Reinforcement Learning (ARL) trains it against a curriculum of benign and malicious image manipulations generated by an adversarial attacker agent. On the CelebA and CelebA-HQ benchmarks, the authors report improvements of over 4.5% and more than 5.3% respectively under challenging manipulation scenarios.

Researchers address deepfake detection, where high-fidelity generators remove explicit traces and leave only subtle frequency-domain artifacts. They propose a Hierarchical Spatial-Frequency Fusion (HSFF) module with frequency attention, paired with a Collaborative Multi-Subband Enhancement (CMSE) mechanism that combines Dual-Tree Complex Wavelet Transform (DTCWT) features with pre-trained visual features across six directional subbands. The method reports 97.2% cross-dataset AUROC on Celeb-DFv2 and 89.1% on DFDC.

Amber (Adaptive Affinity Memorization with Layer Mutation) is a continual learning method for Multimodal Deepfake Continual Detection (MDCD), the task of detecting new multimodal deepfake techniques as they emerge. The authors argue that vanilla continual learning fails here because artifact feature drift, caused by the cross-modal gap, makes preserved memory unrepresentative and drives forgetting of artifact features. Amber combines Memorization Affinity Estimation, which keeps high- and low-affinity memory, with Knowledge Layer-wise Mutation, which broadens historical memory, and is evaluated with a new metric, initial comprehension capability (ICC). The authors report that extensive experiments show Amber adapts to new deepfake techniques while retaining prior knowledge.

Researchers Shuai Wang, Gaobo Yang and Hanling Zhang published a deepfake detection study in the Journal of Information Security and Applications, Volume 101, in September 2026. The source text provided contains only the publication metadata and no description of the method or findings.

Researchers propose the Degradation Consistency Learning Framework to improve deepfake face detection on forgery methods unseen during training. The method exploits degradation inconsistencies between background and manipulated face regions, using a data generation network and a detection network mining both spatial and frequency domain clues, coupled through adversarial training. The authors report effectiveness on benchmark datasets under in-dataset and cross-dataset protocols.

FaceReclaim is a diffusion-based framework that tries to restore the original face hidden inside a face-swapped deepfake image, a task the authors call Traceability of Face-swapping Deepfake (TFD). It reformulates TFD as image-to-image editing and uses Multi-Scale Face Attribute Decoupling and Multi-Modal Face Identity Prompting as dual conditioning. The authors report that synthetic faces retain subtle traces of the original face, and they evaluate restoration both visually and through face verification.

The authors introduce a taxonomy and benchmarking framework that classifies deepfake attacks by intent and generative mechanism, then evaluate them against deep face recognition systems. Diverse deepfake attacks commonly exceed 70% success and, in some regimes, surpass 90%. The study finds that attack success depends on the degree of identity controllability rather than visual quality.

PVLM is a parsing-aware vision-language model with dynamic contrastive learning for zero-shot deepfake attribution, meaning it traces forged faces to generators not seen in training, including diffusion models. The authors build a fine-grained ZS-DFA benchmark and use face parsing to exploit differences in how GAN and diffusion generators preserve source facial attributes. They report that the model exceeds the state of the art on the ZS-DFA benchmark across various protocol evaluations.

Researchers propose a proactive Deepfake detection framework that embeds a watermark into the Quaternion Radial Harmonic Fourier Moments (QRHFMs) domain of face images. A Local-Enhanced State Space (LSS) block extracts features, and a dual-branch architecture pairs a watermark extractor with a forgery discriminator via knowledge distillation. On benchmark datasets the method reports 92.58% average accuracy and 95.96% AUC, with robustness to image quality degradation and to Deepfake techniques unseen during training.

The authors introduce DFREC, a scheme that recovers both the source and target faces from a deepfake image to support identity tracing in forensic investigations. It uses an Identity Segmentation Module, a Source Identity Reconstruction Module and a Target Identity Reconstruction Module, the last of which relies on a Masked Autoencoder. Evaluated on FaceForensics++, CelebaMegaFS, FFHQ-E4S and Celeb-DFv2, DFREC outperforms state-of-the-art deepfake recovery algorithms and is described as the only scheme that recovers both pristine faces with high fidelity.

Researchers propose a deepfake face detection method that recasts detection as a Visual Question Answering task for a Visual Language Model. Component-specific prompts direct attention to the eyes, nose and mouth, and a Q-Former module adjusts visual feature focus according to those prompts. The authors report that the method outperforms existing techniques in detection accuracy and robustness, though the source text gives no numerical results.

Researchers propose a Two-Stage Defense Framework (TSDF) against deepfakes that aims to keep active defenses effective after attackers retrain their models. The framework uses dual-function adversarial perturbations that distort forged results and also poison the data preparation step of an attacker's retraining pipeline. The source reports that traditional interruption methods degrade sharply under adversarial retraining, while TSDF shows stronger dual defense capability.

This ACM Computing Surveys paper, published in Volume 58, Issue 10 (July 2026), examines fairness in deepfake detection across ethnicity, gender and age. The source text provided contains only bibliographic details and no findings, method or results.

FauForensics is a framework for detecting audio-visual deepfakes that uses facial action units, which quantify facial muscle activity, as forgery-resistant features. It computes frame-wise audio-visual similarities through a fusion module with learnable cross-modal queries, aligning lip-audio relationships to address feature heterogeneity. Tested on four public datasets, it reports state-of-the-art results with a 5.17% average cross-dataset improvement over existing methods.

This paper proposes a triple-branch network for deepfake detection that jointly learns spatial features and frequency features from the original image and from images reconstructed through different frequency channels. The authors add feature decoupling and fusion losses derived from mutual information theory to make the model focus on task-relevant features. The method reportedly achieves state-of-the-art performance across six large-scale benchmark datasets.

The paper proposes FoVB (Forgery-aware Audio-Visual Adaptation with Variational Bayes), a framework for multi-modal deepfake detection. It models audio-visual correlation as a Gaussian latent variable estimated via variational Bayes, using difference convolutions and a high-pass filter to extract forgery traces from both modalities, then factorizes the variable with an orthogonality constraint. The authors report that FoVB outperforms other state-of-the-art methods across various benchmarks.

Researchers propose RSG-DA, a framework for deepfake detection that targets poor generalization to unknown images and novel forgery types. It combines a Dynamic Landmark Diffusion Generator that synthesizes hybrid forgery samples, a Dual Data Augmentation strategy, and a Lightweight Generic Forgery Distillation module. Experiments report consistent gains over state-of-the-art methods in intra-dataset and cross-dataset evaluations.