PK's Bio

Talk Zero: Ethical, Responsible, and Safe Requirements for AI
As AI systems grow in capability and adoption, ensuring their safety, robustness, and alignment is more critical than ever. In this talk, I will discuss some of the ongoing existing research at Precog-at-IIITH, in AI safety through the lenses of oversight, interpretability, representation learning, and unlearning. I will discuss the challenges of AI oversight, where one language model evaluates or supervises another and show how model similarity can bias judgments [ICML 2025a]. I will showcase how undesirable behaviors like bias and toxicity in LMs can be mitigated using optimal steering of internal representations [ICML 2024a], and how fixed random projections can surprisingly outperform learned features in continual learning settings [NeurIPS 2024]. On the unlearning front, I will present techniques to remove harmful influences from training data: Cognac for graph data [ICML 2025b], Unlearning [TMLR 2024], and discuss our WMDP benchmark [ICML 2024b].

We are also very excited about making GNNs more interpretable, and will share about FORGE, a framework that captures higher-order node interactions for more faithful explanations [AAAI 2025]. In addition, I will discuss our recent work on evaluating how well large language models adhere to external label definitions [EMNLP 2025a] and on developing semantic-aware foundation models for reasoning over knowledge graphs [EMNLP 2025b]. Together, these works highlight emerging directions to improve the safety, interpretability, and accountability of modern AI systems. ChatGPT was used to edit / polish this Abstract!

[ICML 2025a] oversight risks in aligned models [Link]
[ICML 2024a] Representation Surgery: Theory and Practice of Affine Steering. [Link]
[NeurIPS 2024] Random Representations Outperform Online Continually Learned Representations [Link]
[ICML 2025b] corrective unlearning for graph neural networks [Link]
[TMLR 2024] Corrective Machine Unlearning [Link]
[ICML 2024b] The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning. [Link]
[AAAI 2025] Graph Explanations from Higher Order Structures [Link]
[EMNLP 2025a] LLMs’ adherence to external label definitions [Link]
[EMNLP 2025b] semantic-aware foundation models for knowledge graphs [Link]

Talk one: Data Science for Social Good
Abstract: With increase in sage of the Internet, social media platforms, huge amount of data is getting generated at various level; one of the dire need of the hour is to see how we can use this data to have societal impact. I will briefly mention some super cool projects that we have worked on and that have made some visible contributions to the world outside academia. All work presented in the talk (including datasets, code, slides, recorded videos) has related publications at https://precog.iiit.ac.in/ Projects that I will touch upon are — Selfie deaths / KillFie http://labs.precog.iiitd.edu.in/killfie/ #GeneralElections2019 http://bit.ly/elections19(Mis)Information / Fake content on Twitter, Facebook, and WhatsApp, and Computing for Medicine. We all understand, many real world problems cannot be addressed by a single domain faculty / researcher, we need more students & faculty to come together, hoping to generate some interest to join hands to have an impact and make a difference. Many of our research work is made available for public use through tools or online services at https://precog.iiit.ac.in/ including datasets that we have used in our research work. Our work derives techniques from Computational Social Science, Data Science, Statistics, and Network Science. I will be more than happy to clarify, discuss, any of our work in detail, as required, after the talk.

Talk two: “Privacy and Security in Online Social Media (PSOSM)”
Abstract: With increase in usage of the Internet, there has been an exponential increase in the use of online social media on the Internet. Websites like Facebook, Instagram, Twitter, Tinder, YouTube, and many others have changed the way the Internet is being used. There is a dire need to investigate, measure, and understand privacy and security on online social media from various perspectives (computational, cultural, psychological). Real world scalable systems need to be built to detect and defend security and privacy issues on online social media. I will describe briefly some cool projects that we work on: (Mis)Information / Fake content on Twitter, Facebook, and WhatsApp; Linking profiles across different social networks; Selfie deaths and KillFies; and Manipulated followers. Many of our research work is made available for public use through tools or online services at precog.iiit.ac.in including datasets that we have used in our research work. Our work derives techniques from Computational Social Science, Data Science, Statistics, and Network Science. I will be more than happy to clarify, discuss, any of our work in detail, as required, after the talk.

Talk three: Social Computing for Medicine
Abstract: With increase in importance of computing in healthcare, there has been an exponential increase in use of computing methodologies to develop better understanding of health care systems. In the history of human lives, Health and Medicine have never seen the importance as it has in the current scenario. Perhaps for the first time in the history of democratic India, healthcare is predicted to play an important role in future election manifestos as well. I will briefly describe some cool projects that we are currently working on - WASHKaro http://precog.iiitd.edu.in/research/c4m to provide right information at the right time during pandemic situations like #COVID19; use social media as a source to study mental health & early understanding of psychological patterns of users http://covibes.tavlab.iiitd.edu.in/; build transparent Agent Based Models, classical compartmental models to allow interactions with the assumptions like lockdown, testing, & social interventions. covision http://covision.tavlab.iiitd.edu.in/ a system that we build to showcase these models. The talk will end with showcasing some grand challenges that I see in this space which can be a great starting point for students interested in pursuing research and solving some real world problems. Many of our research work is made available for public use through tools or online services at precog.iiit.ac.in including datasets that we use for research. Our work derives techniques from Social Computing, Computational Social Science, Machine Learning, Statistics, and Network Science. I will be more than happy to clarify, discuss, any of our work in detail, as required, after the talk.

Short Academic Bio [updated in Oct 2024]:

Prof. Ponnurangam Kumaraguru ("PK") is a Professor of Computer Science at IIIT-Hyderabad. He is an Adjunct Faculty at IIT Madras and IIIT Delhi. PK is a Serial Academician with formal appointments (Adjunct/Visiting) in India and across the world, including Brazil, Europe, and USA. He is a senate / academic council member at IIIT Una, JKLU, and LNMIIT Jaipur. PK was inducted an ACM Distinguished Member in 2021. PK is currently (from July 2024) the Vice President of ACM India Council and has been a member since July 2020. PK was listed in the World's 2% Scientists by Stanford University in Nov 2020. PK is a TEDx and an ACM Distinguished & ACM India Eminent Speaker. PK received his Ph.D. from the School of Computer Science at Carnegie Mellon University (CMU). PK's research interests include responsible & safe AI, applied machine learning & NLP, and social network analysis. His Ph.D. thesis work on anti-phishing at CMU contributed in creating an award-winning startup - Wombat Security Technologies, wombatsecurity.com. Wombat was acquired in March 2018 for USD 225 Million. PK started and successfully manages Precog, a research group at IIIT-Hyderabad. Please connect with PK at Twitter and LinkedIn.

PK can be reached at pk.guru@iiit.ac.in