Abstract
Deepfake detection models often struggle with generalization to unseen forgery techniques. This paper investigates this issue through a game-theoretical lens, identifying "toxic" low-order visual interactions as a primary source of poor cross-dataset performance. By decomposing model predictions into multi-order interactions using Shapley values, we demonstrate that low-level patterns often introduce dataset-specific biases. We propose an inference-time strategy to suppress these toxic signals, significantly enhancing detector generalization across various architectures without retraining.
Introduction
The rapid evolution of deepfake technology has created a "cat-and-mouse" game where detectors trained on specific datasets (e.g., FaceForensics++) often fail when encountering novel manipulation techniques (e.g., Celeb-DF). While existing research focuses on data augmentation or architectural tweaks, the underlying cause of this generalization gap remains poorly understood. This work adopts a diagnostic approach to identify the internal representations that hinder a model's ability to generalize.
Methodology: Game-Theoretic Interaction Analysis
We treat deepfake detection as a cooperative game where different image patches (e.g., eyes, nose, mouth) act as players collaborating to produce a final classification score. Using Shapley values, we decompose these contributions into multi-order interactions:
- Low-order interactions: Collaborations between small groups of patches, representing local, simple patterns.
- High-order interactions: Collaborations across large groups of patches, representing global, complex consistency.
We hypothesize that low-order interactions often capture "toxic" dataset-specific artifacts (like specific blur patterns) rather than generalizable features of "fakeness."
Hypotheses and Verification
The study proposes three key hypotheses to explain generalization performance:
- H1: It is difficult to detect genuine images using only fragmentary patches.
- H2: Models with strong generalization encode low-order interactions with fewer negative (toxic) contributions than those with poor generalization.
- H3: Suppressing low-order interaction strength can weaken these negative contributions and improve performance.
To verify these, we developed mathematical metrics ($D^m$ and $\rho^m$). Experimental results across multiple backbones (ResNet-18, ResNet-34, Xception, EfficientNet-b3) confirmed that low-order interactions frequently provide negative contributions to the detection task on unseen data, effectively acting as noise that misleads the model.
Results: Generalization via Toxic Suppression
By visualizing the interactions at different orders, we observe a clear transition: low-order interactions often highlight random background noise or edges, while high-order interactions focus on semantically meaningful facial regions like the eyes and mouth.
We introduced a generic inference-time strategy that subtracts the contribution of "toxic" low-order interactions (typically defined as $m < 0.3L$) from the final prediction. This method requires no retraining and consistently improves cross-dataset performance. For instance, a standard ResNet-18's Frame-AUC on Celeb-DF (v1) improved from 56.76% to 57.56%. The strategy is also compatible with state-of-the-art detectors like SBI and CADDM, showing it can complement existing defense techniques.
Conclusion and Future Work
This research provides a novel game-theoretical framework for diagnosing and improving deepfake detector generalization. By identifying and suppressing "toxic" low-level biases at inference time, we can enhance model robustness without the need for expensive retraining or additional data. Future work will explore the sensitivity of the interaction threshold $m$ and investigate the applicability of this framework to other modalities, such as deepfake audio and video.
Summary updated