PairCLIP‑SWA is a single‑backbone deepfake detector that fine‑tunes an OpenCLIP ViT‑L/14 visual tower using a class‑balanced corpus created by CLIP‑similarity‑guided subsampling and then applies stochastic weight averaging over the fine‑tuning trajectory. Trained on the IEEE SP Cup 2025 (DF‑Wild‑Cup) benchmark, the weight‑averaged model achieves 97.04 % accuracy, 98.64 % AU‑ROC and a 4.07 % equal error rate—outperforming the best published three‑backbone ensemble. The study shows that averaging stabilizes validation performance, prevents misranking on disjoint test sets, and that CLIP pre‑training, rather than ensembling, is the key driver of accuracy.
By Md. Nakib, Sudipto Sarkar, A. J. A. Jubayer Talukder, Mohammad Ariful Haque
arXiv:2512.17730v2 Announce Type: replace
Abstract: Detectors of AI-generated images tend to inherit the biases of the data they are trained on: models fitted to GAN imagery learn to treat GAN-specif...
By Yichen Jiang, Mohammed Talha Alam, Sohail Ahmed Khan, Duc-Tien Dang-Nguyen, Fakhri Karray
arXiv:2610.29152v1 Announce Type: cross
Abstract: A generated chest radiograph can be useful as declared augmentation, but the same image becomes unsupported evidence when submitted as the record of...
By Shaikh Akib Shahriyar, Zinedine Sehili, Adam Arthur, Christopher Schwartz
arXiv:2607. 06615v1 Announce Type: cross Abstract: Image forgery detection is a critical task in digital forensics, yet many deep-learning localization approaches are typically GPU-accelerated and computationally heavier than handcrafted screening methods.
By Sujith K Mandala
arXiv:2607. 13234v1 Announce Type: cross Abstract: Deepfake detectors that achieve near-perfect scores on academic benchmarks collapse on real-world content: recent in-the-wild evaluations report AUC drops of 45-50% for state-of-the-art open-source models.
By Ken Jon Miyachi, Dylan Uys
CropCop is a closed‑set plant‑health recognition system covering 120 operational classes, built from a rigorously audited dataset of 109,107 images after removing 3,233 duplicate relationships. The model, based on a fine‑tuned DINOv3 ConvNeXt‑Tiny, achieves 98.51% accuracy and 96.87% macro‑F1 on a locked internal test, while a quantised MobileNetV4 variant reaches 98.46% accuracy and 96.23% macro‑F1 in a 22.60 MiB runtime artifact. Validation‑only post‑training quantisation and a compact ExecuTorch/XNNPACK PTE ensure high fidelity between the trained model and its deployed form, with minimal decision changes between the INT8 graph and the final artifact.
By Rana Muhammad Ahmed, Sabahat Abbas
BG-REAL is a public benchmark for detecting and localizing background manipulations in images, built from Open Images V7 with 7,000 processed samples across 1,200 source groups. It includes 6,000 real-data anchored samples, 1,000 synthetic controls, six edit families, matched authentic controls, and extensive quality‑control measures. The release offers a construction pipeline, evaluation protocol, and three external baseline results, highlighting false‑positive rates due to re‑encoding artifacts.
By Bugra Alperen Uluirmak, Rifat Kurban
The study evaluates 16 AI‑generated image detection methods in a zero‑shot setting, testing 23 pretrained variants on 12 datasets with 2.6 million images from 291 generators. Results show no single detector consistently outperforms others, with accuracy ranging from 37.5 % to 75.0 %. The study highlights that training data alignment and generator type significantly affect performance, especially against modern commercial models, and identifies systematic failure patterns across datasets.
By Simiao Ren (Neo), Yuchen Zhou (Neo), Xingyu Shen (Neo), Kidus Zewde (Neo), Tommy Duong (Neo), George Huang (Neo), Hatsanai (Neo), Tiangratanakul (Dennis), Tsang (Dennis), Ng (Dennis), En Wei, Jiayu Xue
arXiv:2510.03245v3 Announce Type: replace-cross
Abstract: State-of-the-art attribution methods such as AttEXplore rely on adversarial sample generation that applies an all-pass filter across the freq...
By Ali Yavari, Alireza Mohamadi, Elham Beydaghi, Philipp Seeb\"ock, Rainer A. Leitgeb
The paper evaluates whether multimodal large language models (MLLMs) can serve as forensic detectors of deepfake facial images. Six MLLMs—including GPT-4o, GPT-5, GPT-5.2, Gemini 2.5, Gemini 3, and Seed 1.6—were tested on a dataset of 400 images (50 authentic and 350 synthetic) generated by seven leading image models. Results show that several MLLMs achieve high detection accuracy, with GPT-5.2 reaching 97.9% in pairwise comparisons, but the presence of invisible SynthID watermarks can reduce detection performance by about 10% for Gemini.
By Olivia Salamon, Virginia N. L. Franqueira, Shujun Li
The study challenges the prevailing belief that frequency-based features and compression-invariant learning are essential for robust deepfake detection. Using a controlled, pre‑registered protocol, a simple EfficientNet‑B0 trained on diverse multi‑quality data outperformed the more complex CAFRL model across all compression levels, with a 3.66 AUC point advantage at CRF 40. After identifying and correcting four experimental defects, the authors found that frequency features added no marginal benefit, while data diversity—particularly real constant‑rate‑factor variants—proved to be the key factor for robustness against H.264 re‑encoding.
By Abbas Aliyev, Samir Rustamov
arXiv:2610.34545v1 Announce Type: new
Abstract: Visual-token compression is usually evaluated by clean accuracy and efficiency, but these averages do not reveal failures that compression introduces u...
By Isaac (Rucheng), Li