Threat Landscape of Adversarial Attacks on Generative AI and Large Language Models (LLMs): Exploring Different Types of Adversarial Attacks, Associated Risks, and Mitigation Strategies

Generative Artificial Intelligence (Generative AI) has emerged as a transformative catalyst across disciplines and applications, fundamentally enhancing creativity, productivity, personalization, and problem-solving by creating novel, coherent, and contextually relevant content. Large Language Models (LLMs) are a specific type of generative AI that focuses mainly on understanding, generating, and manipulating human language. The proliferation of LLMs has amplified the threat of adversarial attacks that maliciously manipulate inputs or training data with the adversarial intent to exploit, compromise, or mislead LLMs. Unlike conventional cyberattacks that exploit commonly known software vulnerabilities or directly attack the IT infrastructure, adversarial attacks on LLMs exploit the statistical and linguistic patterns the LLMs have learned. These attack strategies coupled with the sheer scale of LLM deployments and diversity of inputs, exacerbates detection and mitigation challenges of ever-evolving adversarial attacks. Therefore, this paper will explore adversarial attacks on LLMs, and their associated risks and mitigations. Initially, it will explain what adversarial attacks on LLMs are and how they differ from conventional cyberattacks. Next, it will elucidate several types of adversarial attacks on LLMs to provide an in-depth overview of their nature and impact. Afterwards, it will review various risks associated with adversarial attacks on LLMs. Lastly, it will present various mitigation strategies for adversarial attacks on LLMs. This detailed analysis of adversarial attacks on LLMs, along with their associated risks and mitigation strategies, aims to provide in-depth insights into the security and safety challenges inherent to LLM deployment and usage.