S04EP02 | Reversing Large Deep Learning Models | Yashodhan Mandke

Have you ever thought about how an attacker might reverse-engineer an AI model? Our guest today is doing just that, going beyond passwords and keys to unpack the very DNA of deep learning!In this segment, we're diving into the groundbreaking work of reversing large deep learning models. Our guest reveals how it's possible to reverse an AI model's entire mathematical structure, exposing its architecture, critical hyperparameters, and even the internal weights and biases that define its behavior. We'll explore this new frontier of security research in the context of different model formats and major models like GoogleNet and Llama. This isn't just about finding vulnerabilities; it's about understanding how a malicious actor could exploit the sparsity of a tensor or reverse a tokenizer, fundamentally subverting an AI's core logic. This is next-level threat intelligence, showing us how to defend AI by understanding its deepest secrets.Guest: Yashodhan Mandke, Research Scholar MIT-WPUYashodhan is a Security Researcher with over 13 years of cutting-edge experience at the intersection of IoT and AI innovation. A tech visionary currently pursuing a doctorate in Satellite and Security, Yashodhan’s academic journey spans M.Tech in Satellite Communication, M.Tech in Signal Processing, and a B.E. in Electronics & Telecommunication.Recommended reading/viewing, Paper(in this topic) for practitionershttps://goa2025.nullcon.net/doc/goa-2025/nullcon_2025_rev_dl.pdf--Follow us on LinkedIn: https://www.linkedin.com/company/breakpoint-security-podcastAudio on Buzzsprout: https://breakpoint.buzzsprout.comIf you like to see more like this, please Subscribe to Breakpoint Youtube! Please Share with others in the community. It always means a lot!Follow us on LinkedIn: @breakpoint-security-podcastAudio on Buzzsprout: https://breakpoint.buzzsprout.comBuzz me on Twitter or LinkedIn Connect with me on -Twitter: @NeeluTripathy LinkedIn: @neelutripathy