Agility of in Deep Learning Algorithms in Developing Conflicting Attacks Protections and to Secure Vulnerabilities
Renas Rajab Asaad, Saman M. Almufti, Ahmed Alaa Hani, Hawar Bahzad Ahmad, Amira Bibo Sallow, Subhi R. M. Zeebaree · 2024
Deep learning methods have brought revolution to a variety of applications: from image recognition to natural language processing, and many others. However, their most important vulnerability, when used in adversarial conditions, is a matter of concern that remains up to today. Most importantly, adversarial attacks insinuate the design of input in a way to be capable of bringing errors in the model, hence exposing significant vulnerabilities. This paper investigates the nature of such vulnerabilities and proposes the enhancement of model robustness through comprehensive methodologies. First of all, adversaries would be considered under the threat landscape and classified according to their complexity and impact, should one materialize. The second section would detail all available defense mechanisms, again classified based on strengths and limitations. We propose using the ensemble of adversaries, regularization methods, and dynamic network architecture. We validate this using the empirical testing of the technique over several datasets and models, subjected to a battery of rigorous tests.