AI is transforming cybersecurity - not just in cyber defense, but also in how we build software. From detecting threats to writing secure code, AI tools are speeding up development and security operations. But this speed brings risk. Unchecked AI-generated code (“vibe coding”), fast deployments, and apps built by non-tech users are opening new threat vectors. This talk explores how AI is reshaping digital defense and why security must be part of every AI-driven workflow.

Piyush Kohli is a seasoned cybersecurity professional with over 13 years of experience, having worked with some of the industry’s leading organizations including Palo Alto Networks, Cybereason, Trustwave (Singtel/Optus), Deutsche Bank, and HCL. He currently serves as a Senior Security Architect at Palo Alto Networks, where he helps organizations strengthen their defenses against sophisticated cyber threats. Throughout his career, Piyush has been at the forefront of security innovation working with cutting-edge technologies, designing advanced security solutions, and collaborating with global teams to tackle complex cybersecurity challenges.

Ensuring content integrity in livestreaming platforms poses unique challenges compared to short videos, as streams are continuous, multimodal, and require real-time moderation at scale. In this talk, I will present TikTok's approach to livestream deduplication, where long-running broadcasts are segmented into short clips and analyzed through a dual-path system in real-time. The first path employs supervised classification to detect known violation categories, while the second path leverages a large-scale retrieval framework to identify near-duplicate or re-broadcasted content. Candidate clips retrieved from this system are refined using a lightweight multimodal re-ranking model distilled from a stronger teacher model, enabling accurate cross-modal comparisons without sacrificing latency. Finally, clip-level matches are aggregated into session-level decisions to ensure robust enforcement at the room level. This hybrid design, combining retrieval, re-ranking, and knowledge distillation strikes a balance between precision, scalability, and adaptability, making it effective for safeguarding livestream ecosystems against duplication and misuse.

I am a Machine Learning Engineer at ByteDance (TikTok), where I work on advancing video understanding with computer vision and multimodal learning techniques. I received my PhD in Computer Science from the National University of Singapore. My doctoral research centered on analyzing and detecting synthesized face media, with a particular focus on deepfakes.