Security in AI

An intelligent system is a high-value target for new and sophisticated cyber threats. AI security is the critical discipline that protects machine learning models and their data from attacks, manipulation, and theft. This involves defending against adversarial attacks that deceive models, preventing data poisoning, and ensuring the privacy and integrity of the entire AI pipeline.

Key Technologies & Frameworks:

Adversarial Robustness Toolbox (ART): A comprehensive library for evaluating model vulnerabilities to adversarial attacks and implementing defense mechanisms.

Homomorphic Encryption (Microsoft SEAL): Cutting-edge cryptographic techniques that allow computations to be performed on encrypted data, enabling secure “blind” processing.

Model Obfuscation & Watermarking: Techniques to protect intellectual property by making models difficult to reverse-engineer and embedding unique identifiers to trace their origin.

Key Projects:

AI Model Vulnerabilities: Backdoors in LLMs and Beyond: A project that explores the vulnerabilities of AI models, particularly large language models (LLMs), to backdoor attacks and other security threats. It includes techniques for detecting and mitigating these risks, ensuring the integrity and reliability of AI systems.

Security benchmarking for large language models: A presentation that discusses the security challenges and benchmarks for evaluating the robustness of large language Models against adversarial attacks and vulnerabilities.