Tech Talk: Data Challenges in ML-based Security Tasks

The San Francisco Bay Area chapter of the IEEE Computer Society invites to our free and open Virtual Tech Talks (no IEEE membership required):
Speaker: Shravya Kanchi
Title: Data Challenges in ML-based Security Tasks
Abstract: Machine learning classifiers are widely used for security tasks, and efforts to improve them have largely focused on algorithmic advances, while the data challenges that limit their performance have received far less attention. This talk asks a simple question: can Generative AI address these data challenges? We augment training datasets with synthetic data produced by GenAI techniques to improve classifier generalization, evaluating the approach across 7 diverse security tasks using 6 state-of-the-art GenAI methods, and introduce Nimai, a novel scheme that enables highly controlled data synthesis. We find that GenAI can significantly improve security classifiers, with gains of up to 32.6% even in severely data-constrained settings of roughly 180 training samples, and that it enables rapid adaptation to concept drift after deployment with minimal labeling. We also share where it breaks down: some GenAI schemes fail to initialize on certain security tasks, and task characteristics such as noisy labels, overlapping class distributions, and sparse feature vectors hinder the benefit. The talk closes with what these findings mean for building the next generation of GenAI tools for security. This work was accepted at ACM Asia CCS 2026 and is available at (https://www.google.com/url?q=http://arxiv.org/abs/2507.06092&sa=D&source=calendar&ust=1789963223372342&usg=AOvVaw3HOPUVBQUJ6HQ8xxfIrVSC).
Bio: Shravya Kanchi is an Applied AI Scientist at Kai Cyber Inc. She completed her Ph.D. in Computer Science at Virginia Tech, USA, advised by Dr. Daphne Yao. Her research sits at the intersection of security and generative AI, spanning LLM safety, synthetic data generation for security applications, adversarial machine learning, and automated security testing frameworks. Her work has been published at IEEE S&P, ACM Asia CCS, and ACSAC. She previously earned a Masters by Research in Computer Science from IIIT Hyderabad, India and a B.Tech. in Computer Science from IIIT Sri City, India. She works hands-on with LLMs, GANs, and diffusion models applied to security-specific challenges, with an emphasis on practical defenses and evaluation-driven systems.
Speaker(s): Shravya Kanchi
Virtual: https://events.vtools.ieee.org/m/577937