Enhancing Trustworthy Deep Learning for Image Classification against Evasion Attacks: A systematic literature review

Abstract Deep learning (DL) algorithms have demonstrated remarkable performance in image classification , making them essential for critical applications. However, recent research has highlighted the vulnerability of DL to cyber-attacks, such as data poisoning and evasion attacks, which can significantly degrade its performance. Addressing this issue and developing trustworthy and robust DL models against such attacks remain ongoing research challenges. This study aims to evaluate the state-of-the-art Adversarial Machine Learning techniques (AML) in image classification through a systematic literature review encompassing 49 different studies conducted over the last decade. By conducting this review, potential obstacles in building robust DL models capable of withstanding evasion attacks are identified. The comprehensive analysis of existing literature is crucial to explore the impact of adversarial attacks on the security and resilience of DL models , which aligns with the regulations set forth by the European Commission’s AI High-Level Expert Group (HLEG) to ensure the reliability of artificial intelligence. Furthermore, this review examines the key features of adversarial attacks and proposes a reliable and trustworthy framework based on AML. This framework aims to effectively mitigate previously unknown attacks, resulting in reliable and explainable predictions.