LAA-Net: Localized Artifact Attention Network for Quality-Agnostic and Generalizable Deepfake Detection

TL;DR

LAA-Net: Localized Artifact Attention Network achieves 99.96% AUC in high-quality deepfake detection by explicitly focusing on local artifacts using multi-task learning and enhanced feature pyramid networks.

cs.CV 🔴 Advanced 2024-01-25 136 citations 55 views
Dat Nguyen Nesryne Mejri Inder Pal Singh Polina Kuleshova Marcella Astrid Anis Kacem Enjie Ghorbel Djamila Aouada
deepfake detection attention mechanism multi-task learning feature pyramid network local artifacts

Key Findings

Methodology

This paper introduces LAA-Net, which combines explicit attention mechanisms with multi-task learning to detect high-quality deepfakes. The model employs three parallel branches: a binary classifier, a heatmap regression, and a self-consistency regression, trained on synthetically generated pseudo-fake data derived from blending techniques. The heatmap branch localizes vulnerable pixels likely to contain blending artifacts, while the self-consistency branch measures the similarity of pixels relative to randomly selected vulnerable points. The core innovation is the integration of these branches within a unified framework, guiding the network to focus on subtle local artifacts. Additionally, the proposed Enhanced Feature Pyramid Network (E-FPN) propagates low-level discriminative features across scales, reducing redundancy and enhancing the model's sensitivity to fine-grained cues. The training process involves joint optimization with a composite loss function, leveraging data augmentation and pseudo-fake synthesis, enabling the model to generalize well across unseen datasets.

Key Results

  • On the FF++ dataset, LAA-Net combined with Self-Blended Images (SBI) achieves an AUC of 99.96%, outperforming existing methods such as Multi-attentional and RECCE by over 27%.
  • Cross-dataset evaluations on Celeb-DFv2, DFD, DFDC, and DFW show that LAA-Net maintains high performance, with average AUC improvements exceeding 15%, demonstrating strong generalization to different fake generation techniques and quality levels.
  • Ablation studies confirm that the explicit attention modules and E-FPN contribute approximately 10-15% performance gains individually, with their combination yielding the best results, validating the effectiveness of the proposed architecture.

Significance

This work addresses the critical challenge of detecting high-quality deepfakes that contain subtle, localized artifacts. By explicitly modeling these artifacts through a multi-task attention framework and multi-scale feature propagation, the proposed LAA-Net significantly enhances robustness and generalization. Its ability to detect unseen manipulations with high accuracy has profound implications for media verification, content authenticity, and digital forensics. The approach sets a new benchmark in deepfake detection, bridging the gap between research and real-world application, especially in scenarios requiring high reliability and low false positives.

Technical Contribution

The paper's main technical contributions include the design of an explicit attention mechanism that combines heatmap localization and self-consistency estimation, enabling the model to focus on small, artifact-prone regions. The introduction of the E-FPN enhances the flow of low-level features across scales, reducing redundancy and improving the discrimination of subtle artifacts. The multi-task learning framework, trained solely on real data augmented with synthetic pseudo-fakes, achieves high generalization performance. These innovations collectively improve detection accuracy, especially for high-quality deepfakes, and open new avenues for fine-grained, robust forensic analysis.

Novelty

This study is the first to explicitly incorporate a dual attention mechanism—heatmap-based localization and self-consistency—within a multi-task learning framework for deepfake detection. The E-FPN architecture further distinguishes itself by effectively propagating low-level features while minimizing redundancy, a significant advancement over traditional FPN variants. Unlike prior works that rely solely on global features or implicit attention modules, LAA-Net explicitly targets local artifacts, leading to superior detection performance and robustness across diverse datasets and manipulation techniques.

Limitations

  • While highly effective, the model's performance may degrade under extreme conditions such as severe noise, occlusion, or very low lighting, which obscure subtle artifacts.
  • The reliance on synthetic pseudo-fake data for training, although beneficial, may not fully capture the diversity of real-world manipulations, potentially limiting detection in novel scenarios.
  • The computational complexity of multi-scale feature fusion and multi-task training demands significant hardware resources, which could hinder deployment in resource-constrained environments.

Future Work

Future research will focus on reducing model complexity to enable real-time deployment, possibly through model compression or pruning. Incorporating temporal and multimodal cues, such as audio-visual inconsistencies, could further improve detection robustness. Exploring semi-supervised or unsupervised learning strategies to reduce dependence on synthetic data and enhance adaptability to unseen fake generation methods is another promising direction. Additionally, extending the framework to video-level analysis and integrating explainability modules will make the system more transparent and trustworthy for practical use.

AI Executive Summary

The rapid advancement of generative AI models has led to an explosion of highly realistic deepfake content, posing serious threats to information integrity, privacy, and security. Traditional detection methods, primarily based on global feature extraction using models like XceptionNet, struggle to identify subtle, localized artifacts embedded in high-quality deepfakes. These artifacts often escape coarse analysis, especially when the fake content is meticulously crafted to mimic genuine signals. Consequently, there is an urgent need for more refined detection techniques capable of focusing on the minute details that betray fake content.

In response to this challenge, the authors propose LAA-Net, a novel deepfake detection framework that explicitly models local artifacts through a multi-task attention mechanism. The core idea is to guide the network's focus toward small, artifact-prone regions—referred to as vulnerable points—by leveraging heatmap and self-consistency branches. These branches are trained on synthetically generated pseudo-fake data, created via blending techniques that simulate real-world manipulations without requiring manual annotation. The heatmap branch localizes potential artifact regions, while the self-consistency branch measures the similarity of pixels relative to randomly selected vulnerable points, reinforcing the model’s attention to subtle inconsistencies.

Complementing this attention mechanism, the authors introduce an Enhanced Feature Pyramid Network (E-FPN), designed to propagate discriminative low-level features across multiple scales efficiently. Unlike traditional FPNs, E-FPN reduces feature redundancy and enhances the flow of fine-grained details, which are crucial for detecting subtle artifacts. This architecture enables the model to leverage both global context and local cues, resulting in a robust detection system.

Extensive experiments on benchmark datasets such as FF++, Celeb-DFv2, DFD, and DFDC demonstrate the effectiveness of LAA-Net. The model achieves an AUC of 99.96% on FF++, surpassing existing methods by a significant margin. Cross-dataset evaluations confirm its strong generalization capabilities, maintaining high performance across different fake generation techniques and quality levels. Ablation studies further validate the contributions of each component, showing that the explicit attention modules and E-FPN collectively boost detection accuracy.

This work marks a substantial step forward in deepfake forensic research. By explicitly modeling local artifacts and optimizing multi-scale feature propagation, LAA-Net addresses the core limitations of prior approaches. Its ability to detect high-quality, unseen deepfakes with high accuracy has profound implications for digital content verification, media security, and legal investigations. While challenges remain—such as computational cost and robustness under extreme conditions—the proposed framework lays a solid foundation for future innovations. Overall, this research enhances our capacity to safeguard truthfulness in the digital age, reinforcing the integrity of visual information in an era of sophisticated synthetic media.

Deep Dive

Abstract

This paper introduces a novel approach for high-quality deepfake detection called Localized Artifact Attention Network (LAA-Net). Existing methods for high-quality deepfake detection are mainly based on a supervised binary classifier coupled with an implicit attention mechanism. As a result, they do not generalize well to unseen manipulations. To handle this issue, two main contributions are made. First, an explicit attention mechanism within a multi-task learning framework is proposed. By combining heatmap-based and self-consistency attention strategies, LAA-Net is forced to focus on a few small artifact-prone vulnerable regions. Second, an Enhanced Feature Pyramid Network (E-FPN) is proposed as a simple and effective mechanism for spreading discriminative low-level features into the final feature output, with the advantage of limiting redundancy. Experiments performed on several benchmarks show the superiority of our approach in terms of Area Under the Curve (AUC) and Average Precision (AP). The code is available at https://github.com/10Ring/LAA-Net.

cs.CV

References (20)

Dynamic Graph Learning with Content-guided Spatial-Frequency Relation Reasoning for Deepfake Detection

Yuan Wang, Kun Yu, Chen Chen et al.

2023 179 citations ⭐ Influential

Learning Self-Consistency for Deepfake Detection

Tianchen Zhao, Xiang Xu, Mingze Xu et al.

2020 397 citations ⭐ Influential View Analysis →

Feature Pyramid Networks for Object Detection

Tsung-Yi Lin, Piotr Dollár, Ross B. Girshick et al.

2016 28046 citations ⭐ Influential View Analysis →

FaceForensics++: Learning to Detect Manipulated Facial Images

Andreas Rössler, D. Cozzolino, L. Verdoliva et al.

2019 3296 citations ⭐ Influential View Analysis →

Focal Loss for Dense Object Detection

Tsung-Yi Lin, Priya Goyal, Ross B. Girshick et al.

2017 33913 citations ⭐ Influential

Detecting Deepfakes with Self-Blended Images

Kaede Shiohara, T. Yamasaki

2022 628 citations ⭐ Influential View Analysis →

Implicit Identity Leakage: The Stumbling Block to Improving Deepfake Detection Generalization

S. Dong, Jin Wang, Renhe Ji et al.

2022 253 citations ⭐ Influential View Analysis →

Xception: Deep Learning with Depthwise Separable Convolutions

François Chollet

2016 18368 citations View Analysis →

Feature Pyramid Network for Multi-class Land Segmentation

Selim S. Seferbekov, V. Iglovikov, A. Buslaev et al.

2018 166 citations View Analysis →

Face2Face: Real-Time Face Capture and Reenactment of RGB Videos

Justus Thies, M. Zollhöfer, M. Stamminger et al.

2016 2298 citations View Analysis →

Grad-CAM: Why did you say that? Visual Explanations from Deep Networks via Gradient-based Localization

Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam et al.

2016 1714 citations

Dlib-ml: A Machine Learning Toolkit

Davis E. King

2009 3603 citations

Capsule-forensics: Using Capsule Networks to Detect Forged Images and Videos

H. Nguyen, J. Yamagishi, I. Echizen

2018 812 citations View Analysis →

Exploiting Visual Artifacts to Expose Deepfakes and Face Manipulations

Falko Matern, C. Riess, M. Stamminger

2019 768 citations

Regularizing Deep Neural Networks by Enhancing Diversity in Feature Extraction

B. Ayinde, T. Inanc, J. Zurada

2019 101 citations

FCOS: Fully Convolutional One-Stage Object Detection

Zhi Tian, Chunhua Shen, Hao Chen et al.

2019 6389 citations View Analysis →

When Does Label Smoothing Help?

Rafael Müller, Simon Kornblith, Geoffrey E. Hinton

2019 2473 citations View Analysis →

Deepfakes

Erhard Taverna

2019 270 citations

FakeSpotter: A Simple Baseline for Spotting AI-Synthesized Fake Faces

Run Wang, L. Ma, Felix Juefei-Xu et al.

2019 61 citations View Analysis →

Deep Learning for Deepfakes Creation and Detection

T. Nguyen, C. Nguyen, Dung Nguyen et al.

2019 151 citations

Cited By (20)

FakeI2V-Bench: Benchmarking the Applicability of Image-level Deepfake Detectors for Deepfake Video Detection

2026 ⭐ Influential View Analysis →

LOGER: Local-Global Ensemble for Robust Deepfake Detection in the Wild

2026 1 citations ⭐ Influential View Analysis →

LAA-X: Unified Localized Artifact Attention for Quality-Agnostic and Generalizable Face Forgery Detection

2026 ⭐ Influential View Analysis →

ExposeAnyone: Personalized Audio-to-Expression Diffusion Models Are Robust Zero-Shot Face Forgery Detectors

2026 1 citations ⭐ Influential View Analysis →

Toward Robust Proactive Deepfake Detection via Orthogonal Moment Watermarking

2026 1 citations ⭐ Influential

Robust Deepfake Detection, NTIRE 2026 Challenge: Report

2026 24 citations ⭐ Influential View Analysis →

When AUC Misleads: Polarization-Aware Evaluation of Deepfake Detectors under Domain Shift

2026 ⭐ Influential View Analysis →

Low-rank Orthogonal Subspace Intervention for Generalizable Face Forgery Detection

2026 1 citations ⭐ Influential View Analysis →

Constructing adaptive spatial-frequency interactive network with bi-directional adapter for generalizable face forgery detection

2025

Fusion-SSAT: Unleashing the Potential of Self-supervised Auxiliary Task by Feature Fusion for Generalized Deepfake Detection

Beyond Flicker: Detecting Kinematic Inconsistencies for Generalizable Deepfake Video Detection

2025 1 citations View Analysis →

Deepfake Detection Leveraging Self-Blended Artifacts Guided by Facial Embedding Discrepancy

2025 5 citations

On the Holistic Approach for Detecting Human Image Forgery

2026 7 citations View Analysis →

Your One-Stop Solution for AI-Generated Video Detection

2026 8 citations View Analysis →

A Multi-Grained Parallel Spatio-Temporal Learning Architecture for Deepfake Video Detection

2026 1 citations

Beauty and the Beast: Imperceptible Perturbations Against Diffusion-Based Face Swapping via Directional Attribute Editing

2026 1 citations View Analysis →

VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning

2026 12 citations View Analysis →

SimLBR: Learning to Detect Fake Images by Learning to Detect Real Images

Beyond Static Artifacts: A Forensic Benchmark for Video Deepfake Reasoning in Vision Language Models

Multi-view detection of two-wheeled vehicles on urban roads: a novel dataset and a tailored detector

2026