Core Concepts
The proposed architecture injects identity information into a semantic face image synthesis model to improve identity preservation during generation, and exploits this capability to perform inconspicuous adversarial attacks on face recognition systems.
Abstract
The paper presents a novel Semantic Image Synthesis (SIS) method that incorporates identity information during the generation process. The key idea is to extract an identity embedding from a pre-trained face recognition model and use it as an additional style code, which is then injected into the generator through a cross-attention mechanism.
The authors show that this identity injection has two main effects:
Identity Preservation: When using the same identity as the input face, the proposed method significantly improves the preservation of the original identity in the generated face, outperforming state-of-the-art SIS approaches.
Adversarial Attacks: When swapping the identity embedding with that of a different individual, the model can perform inconspicuous adversarial attacks on face recognition systems. The generated face will visually resemble the original subject, but will be recognized by the FR system as the target identity.
The authors conduct extensive experiments to validate these two capabilities. For identity preservation, they report improved cosine similarity scores between the original and generated faces across multiple face recognition models. For the adversarial attacks, they achieve state-of-the-art Attack Success Rates, while maintaining low perceptual differences between the original and attacked faces.
The paper also explores the effect of swapping different facial attributes (e.g., eyes, eyebrows, mouth) on the adversarial attack performance, finding that targeting identity-related regions leads to the most effective and inconspicuous attacks.
Overall, the proposed architecture demonstrates the ability to both preserve identity during semantic face generation and leverage this capability for powerful yet stealthy adversarial attacks on face recognition systems.
Stats
The paper does not provide any specific numerical data or statistics in the main text. The key results are reported in the form of quantitative metrics, such as cosine similarity for identity preservation and Attack Success Rate for the adversarial attacks.
Quotes
"Whereas most systems reached excellent visual generation quality, they still face difficulties in preserving the identity of the starting input subject."
"Preserving the perceived identity is crucial to make synthetic data exploitable in biometrics applications."
"By exploiting the versatility of cross-attentions, we are able to condition the image generation with high-level information such as the identity, in addition to low-level style features, ultimately improving the identity similarity with respect to the input face."