Securing Facial Bioinformation by Eliminating Adversarial Perturbations

Review Pharmacology & Pharmacy

The challenges of generalizability in artificial intelligence for ADME/Tox endpoint and activity prediction

David Z. Huang et al.

Summary: Artificial intelligence (AI) has made significant advancements in various fields such as computer vision, natural language processing, and games. Developing robust and accurate AI models for ADME/Tox endpoint and activity prediction in drug discovery could revolutionize the process. Despite successful applications, a key challenge is determining how generalizable these predictive models are.

EXPERT OPINION ON DRUG DISCOVERY (2021)

添加到收藏夹

Article Computer Science, Theory & Methods

Adversarial Machine Learning Attacks and Defense Methods in the Cyber Security Domain

Ishai Rosenberg et al.

Summary: This article presents a comprehensive summary of recent research on adversarial attacks against security solutions based on machine learning techniques, highlighting the associated risks. The methods of adversarial attacks are characterized based on occurrence stage, attacker goals, and capabilities, while categorizing the applications of attack and defense methods in the cyber security domain. It also discusses the impact of recent progress in adversarial learning fields on future research directions in cyber security.

ACM COMPUTING SURVEYS (2021)

添加到收藏夹

Article Computer Science, Information Systems

Anti-Forensics for Face Swapping Videos via Adversarial Training

Feng Ding et al.

Summary: The study shows that DeepFake technology may pose a potential threat, so researchers are working on developing anti-forensics methods. The GAN model proposed in this paper can effectively combat DeepFake forensics detectors, generate high-quality anti-forensics videos, significantly boosting the level of DeepFake anti-forensics.

IEEE TRANSACTIONS ON MULTIMEDIA (2021)

添加到收藏夹

Proceedings Paper Computer Science, Artificial Intelligence

CAP-GAN: Towards Adversarial Robustness with Cycle-consistent Attentional Purification

Mingu Kang et al.

Summary: Adversarial attacks aim to fool target classifiers with imperceptible perturbations, but the CAP-GAN purification model uses pixel-level and feature-level consistency for effective defense. By incorporating guided attention and knowledge distillation, CAP-GAN outperforms other pre-processing defenses on the CIFAR-10 dataset.

2021 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN) (2021)

添加到收藏夹

Article Engineering, Electrical & Electronic