Gradient Alignment for Cross-Domain Face Anti-Spoofing

TL;DR

GAC-FAS is introduced, a novel learning objective that encourages the model to converge towards an optimal flat minimum without necessitating additional learning modules, and specifically guides the model to be robust against domain shifts.

Abstract

Recent advancements in domain generalization (DG) for face anti-spoofing (FAS) have garnered considerable attention. Traditional methods have focused on designing learning objectives and additional modules to isolate domain-specific features while retaining domain-invariant characteristics in their representations. However, such approaches often lack guarantees of consistent maintenance of domain-invariant features or the complete removal of domain-specific features. Furthermore, most prior works of DG for FAS do not ensure convergence to a local flat minimum, which has been shown to be advantageous for DG. In this paper, we introduce GAC-FAS, a novel learning objective that encourages the model to converge towards an optimal flat minimum without necessitating additional learning modules. Unlike conventional sharpness-aware minimizers, GAC-FAS identifies ascending points for each domain and regulates the generalization gradient updates at these points to align coherently with empirical risk minimization (ERM) gradient updates. This unique approach specifically guides the model to be robust against domain shifts. We demonstrate the efficacy of GAC-FAS through rigorous testing on challenging cross-domain FAS datasets, where it establishes state-of-the-art performance. The code is available at https://github.com/leminhbinh0209/CVPR24-FAS.

Figures (8)

• Figure 1: Illustration of Our Learning Objective. Most SoTA methods zhou2023iadgwang2022ssanliu2023udgfasssdg for DG in FAS rely on auxiliary modules to learn domain-invariant features, and do not guarantee convergence towards a flat minimum. In contrast, our method coherently aligns the generalization gradients at ascending points of each domain with gradients derived from ERM. This approach ensures that the model converges to an optimal flat minimum and is robust against domain shifts.
• Figure 2: Illustration of Different SAM Objective Approach. (a) Standard SAM variants applied to the entire source dataset yield a biased ascending vector $\hat{\epsilon}_{\boldsymbol{\mathcal{S}}}$, predominantly influenced by a particular domain. (b) Domain-specific gradient adjustments lead to noisy gradient estimates, impeding optimization progress. (c) Our proposed GAC-FAS addresses these issues by computing perturbation losses across the dataset at all ascending points, while concurrently adjusting gradients to align with the ERM gradients at the current point (with $\gamma$), making the model robust to domain shift.
• Figure 3: Illustration of the effects described by Eq. \ref{['eqn:sum_sum']}. The term $-\eta\nabla \mathcal{L}_{\text{p}_1}(\mathcal{S}_2)$ represents the generalization gradient update of the model learned from $\mathcal{S}_1$ and $\mathcal{S}_2$. The term $-\eta\nabla\mathcal{L} (\mathcal{S}_3)$ denotes the ERM update for domain $\mathcal{S}_3$ and serves as a comparative oracle for domain shift. In the absence of our regularization, the generalization update is not robust to the domain shift associated with $\mathcal{S}_3$ (left), as their update directions are different. Conversely, with our regularization as formulated in Eq. \ref{['eqn:sum_sum']}, the generalization update aligns with the ERM update on $\mathcal{S}_3$, suggesting that the model updates in a direction that is robust to domain shifts (right).
• Figure 4: Unseen 2D attack
• Figure 5: Ablation study: Sensitivity analysis of hyper-parameters $\gamma$ and $\rho$ on ICM $\rightarrow$ O upon convergence performance.