Unravelling Robustness of Deep Learning based Face Recognition Against Adversarial Attacks (1803.00401v1)

Abstract: Deep neural network (DNN) architecture based models have high expressive power and learning capacity. However, they are essentially a black box method since it is not easy to mathematically formulate the functions that are learned within its many layers of representation. Realizing this, many researchers have started to design methods to exploit the drawbacks of deep learning based algorithms questioning their robustness and exposing their singularities. In this paper, we attempt to unravel three aspects related to the robustness of DNNs for face recognition: (i) assessing the impact of deep architectures for face recognition in terms of vulnerabilities to attacks inspired by commonly observed distortions in the real world that are well handled by shallow learning methods along with learning based adversaries; (ii) detecting the singularities by characterizing abnormal filter response behavior in the hidden layers of deep networks; and (iii) making corrections to the processing pipeline to alleviate the problem. Our experimental evaluation using multiple open-source DNN-based face recognition networks, including OpenFace and VGG-Face, and two publicly available databases (MEDS and PaSC) demonstrates that the performance of deep learning based face recognition algorithms can suffer greatly in the presence of such distortions. The proposed method is also compared with existing detection algorithms and the results show that it is able to detect the attacks with very high accuracy by suitably designing a classifier using the response of the hidden layers in the network. Finally, we present several effective countermeasures to mitigate the impact of adversarial attacks and improve the overall robustness of DNN-based face recognition.

• The paper experimentally evaluates the impact of five adversarial attacks on DNN-based face recognition systems, demonstrating significant performance degradation.
• The research proposes analyzing abnormal filter responses in intermediate DNN layers using Canberra distance to effectively detect adversarial images.
• A selective dropout mechanism coupled with denoising is proposed to help restore recognition performance, paving the way for more resilient systems.

An In-Depth Analysis of Robustness in Deep Learning-Based Face Recognition Against Adversarial Attacks

The paper "Unravelling Robustness of Deep Learning based Face Recognition Against Adversarial Attacks" by Gaurav Goswami et al. presents a comprehensive investigation into the vulnerabilities of deep neural networks (DNNs) employed for face recognition when exposed to adversarial attacks. With the increasing deployment of face recognition systems that utilize DNNs, understanding and mitigating adversarial threats is of paramount importance in ensuring the reliability of these systems.

The authors identify three primary aims: evaluating the robustness of DNN-based models against adversarial attacks, detecting these vulnerabilities through network response analysis, and suggesting practical mitigation strategies. The methodology includes experimental assessments on popular networks such as OpenFace and VGG-Face, considering a range of adversarial distortions simulating real-world scenarios, and two notable databases—MEDS and PaSC.

Key Findings and Methodology

1. Adversarial Attack Impact: By introducing five distinct types of image-level and face-level adversarial distortions—such as grid-based occlusion and noise based on most significant bit alteration—the paper delineates how these distortions result in significant performance degradation. Through rigorous testing, it is demonstrated that deep learning-based face recognition systems, while exhibiting superior baseline performance, are notably more vulnerable than traditional non-deep learning systems to adversarial inputs. Notably, OpenFace and VGG-Face exhibited marked reduction in genuine accept rate (GAR) under adversarial conditions.
2. Detection of Adversarial Distortions: The research proposes a novel method to identify adversarial images by analyzing the abnormal filter response behavior within the intermediate layers of a DNN. This detection mechanism, leveraging the Canberra distance metric, effectively distinguishes between distorted and undistorted images alone using internal layer activations. The results indicate high detection accuracy, especially when applied to the heavily distorted PaSC database.
3. Mitigation Strategies: In response to the detection of adversarial attacks, the authors propose a selective dropout mechanism coupled with denoising techniques. By disabling the most impacted filters of the network identified during detection, the method shows promise in restoring the recognition performance to near-baseline levels. The mitigation effectiveness was validated across multiple distorted datasets using different network architectures.

Implications and Future Directions

This work significantly contributes to the domain of artificial intelligence by advancing the understanding of DNN vulnerabilities in face recognition systems. The experimental framework provides a critical benchmark for evaluating not only existing models but also for testing future architectures against adversarial attacks. Furthermore, the threefold approach encompassing detection, evaluation, and mitigation paves the way towards developing more resilient and secure face recognition systems.

In terms of future work, the paper highlights the importance of integrating complex mitigation frameworks and exploring additional adversarial scenarios to further enhance the robustness of DNNs. The continuation of refining detection algorithms and incorporating adversarial correction mechanisms at runtime will ensure efficacy in dynamic, unpredictable environments.

In conclusion, the research by Goswami et al. underscores the need to prioritize robustness in deploying deep learning systems for critical applications, such as biometric verification, where security and accuracy are paramount. This foundational work is vital for safely harnessing the capabilities of deep learning while safeguarding against adversarial exploitation.