Can Ignoring Identity Features Boost Deepfake Detection? A New Approach

Introduction

With the rapid rise of AIGC‑driven face‑editing tools, creating realistic forged faces has become easier, leading to malicious uses such as fake news and prank videos. Existing deepfake detectors achieve high accuracy on known forgeries but struggle when encountering novel manipulation algorithms.

Problem: Implicit Identity Leakage

Current two‑class detection models inadvertently learn identity features present in the training data. During training, a decision boundary forms based on these identity cues, causing the model to rely on who is in the image rather than on the forgery artifacts. This phenomenon is termed "implicit identity leakage" and harms generalization to unseen forgeries.

Proposed Solution: Identity‑Agnostic Deepfake Detector

The new detector consists of two modules:

Local Forgery Region Detection : A lightweight head added after the backbone that predicts whether each spatial anchor contains forged artifacts, thereby forcing the model to ignore global identity information.

Multi‑Scale Facial Forgery Generation : A data‑augmentation pipeline that creates forged images with annotated fake regions using sliding windows at multiple scales and two fusion strategies (global swap and local swap). This supplies the detector with diverse, region‑labeled training samples.

By training only on these local cues, the model learns common forgery patterns without being misled by identity features.

Experiments

4.1 Verifying Implicit Identity Leakage

Linear identity classifiers were attached to pretrained deepfake detectors and trained on Celeb‑DF, FF++, and LFW. Classification accuracy >50% confirmed that the feature space retained identity information, supporting the leakage hypothesis.

4.2 Backbone Comparison

Various backbones (e.g., ResNet, EfficientNet) were evaluated. Across both same‑domain and cross‑domain tests, the proposed method consistently outperformed baseline two‑class detectors, demonstrating robustness regardless of backbone choice.

Feature Visualization

t‑SNE plots of high‑dimensional features extracted from 100 random samples (10 IDs) show that baseline models separate identities in feature space, while the proposed method produces overlapping clusters, indicating reduced identity reliance.

Results

When identity cues are suppressed, even a simple binary classifier surpasses state‑of‑the‑art industry algorithms on unseen datasets. The detector also offers better interpretability by highlighting forged regions.

Conclusion and Outlook

The study reveals that traditional deepfake detectors suffer from implicit identity leakage, limiting their generalization. By ignoring identity features and concentrating on local forgery artifacts, the proposed algorithm achieves superior cross‑domain performance and provides a promising direction for future research and real‑world deployment.

Figures