Evaluating Adversarial Robustness in Deep Learning–Based Face Recognition with Integrated Liveness Detection

Sammanfattning

Introduction: Deep learning-based face recognition systems are widely used in biometric authentication because they provide accurate and convenient identity verification. However, these systems remain vulnerable to adversarial attacks, where small input perturbations can influence model predictions. Although liveness detection is commonly used to reduce presentation attacks, its effect on adversarial robustness within an integrated authentication pipeline remains unclear. Research Question: This thesis investigates how the integration of liveness detection with face recognition affects robustness against adversarial attacks compared to evaluating the two components separately. Method: A quantitative experimental approach was used. A fine-tuned ArcFace-based face recognition model and a fine-tuned liveness detection model were evaluated independently and as part of an integrated biometric authentication system. The Labeled Faces in the Wild (LFW) dataset was used for face recognition, while the liveness detection model was trained using real facial images from LFW and fake samples extracted from Replay Attack videos. Adversarial examples were generated using the Fast Gradient Sign Method (FGSM) with different perturbation strengths. Performance was evaluated using accuracy, false acceptance rate (FAR), false rejection rate (FRR), and similarity scores. Results: The results show that adversarial perturbations affect the evaluated components differently. The fine-tuned face recognition model exhibited gradual performance degradation as perturbation strength increased, while the liveness detection model maintained relatively stable performance under moderate perturbation levels. Under clean conditions, the integrated system achieved lower accuracy and higher false rejection behavior than standalone face recognition due to stricter authentication decisions. Under moderate adversarial conditions, the integrated system showed smaller changes in some performance measures compared with standalone face recognition. However, under stronger perturbations, the integrated system also experienced substantial performance degradation and increased false rejection behavior. Discussion: The findings demonstrate that integrating liveness detection influences authentication behavior within biometric systems but does not guarantee robustness against adversarial perturbations. Although the integrated system showed smaller performance changes under moderate perturbation levels, the system remained vulnerable under stronger attacks. These findings highlight the importance of evaluating complete biometric authentication pipelines rather than assessing individual components independently. Future work should focus on evaluating stronger attack scenarios, improving adversarial robustness, and investigating more realistic deployment environments.