Ankush Gupta
Adversarial AI Defense in Large Language Models & Generative AI - CyberSecurity Strategy & Challenges
Abstract:
The rapid growth of Large Language Models (LLMs) and generative AI models has enabled a suite of novel capabilities, including advanced analysis of natural language, code synthesis, creative content creation, and multimodal reasoning. Yet, this power brings with it substantial vulnerability to adversarial control. Unlike narrow models of traditional machine learning, generative AI models operate in the open with data from many different sources, other data sources and autonomous applications, and the adversarial surface is increased from text to image/multimodal. New threats appear in three main guises: immediate injection, where malign instructions overwhelm or hijack the running of tools even in safe environments; adversarial examples, where crafted perturbations of inputs corrupt decoding paths to produce harmful, biased, or simply incorrect outputs; and data poisoning, where perturbation of training, fine-tuning, or retrieval corpora introduces latent backdoors, obliterates alignment, or destroys factuality at scale.
Profile:
Ankush Gupta is a Senior Solution Architect at T-Mobile USA Inc. (as a customer of HCLTech) and accomplished researcher and leader with over 21 years of global experience delivering cutting-edge enterprise solutions across telecom, retail, and fintech. Ankush Gupta work sits at the intersection of advanced technology and business impact, combining Cybersecurity, AI/Gen AI, cloud-native systems (AWS, Azure), and intelligent automation to design scalable platforms. Ankush is a Senior member of IEEE, SigmaXi and ACM and has delivered several sessions on Cybersecurity and Artificial Intelligence. He has contributed well as a researcher and scholar on these areas as subject matter expert.