Reproduce “CausalCLIP: Causally-Informed Feature Disentanglement and Filtering for Generalizable Detection of Generated Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bo Liu, Qiao Qin, Qinghui He
CausalCLIP disentangles causal forensic cues from spurious artifacts within CLIP embeddings to improve generalizable detection of AI-generated images. The framework achieves superior cross-model generalization, demonstrating average improvements of 2.32% AP and 4.62% ACC in diffusion-sources evaluation, and 1.23% AP and 3.26% ACC in GAN-sources evaluation, while also enhancing robustness to image perturbations.
Key insights
- Explicitly disentangling stable, causal forensic cues from spurious, generator-specific artifacts is crucial for developing image detectors that generalize robustly across diverse generative models.
- Causal inference principles, including structural causal models and counterfactual interventions, can effectively guide feature learning to isolate and preserve features intrinsically linked to the 'generated' attribute.
- Combining a differentiable feature masking mechanism with adversarial training and statistical independence regularization (HSIC) enables a classifier to reliably depend on robust causal features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.