Hi, I'm Aman!
I'm an AI Researcher at Cisco specializing in AI safety, security, and privacy leakages in AI systems. In my brief time as a researcher, I've been fortunate to publish in various AI conferences, journals, and workshops, with my work spanning privacy-preserving machine learning, AI security, and large language models. My focus has been on uncovering vulnerabilities in foundation models - work that has garnered media attention (1, 2, 3) and led to invitations to join some really cool security initiatives like OpenAI's Red Teaming Network and Anthropic's Model Safety Bug Bounty Program (though couldn't participate completely due to student-visa restrictions).
With a Masters in Privacy Engineering from Carnegie Mellon University, I've worked closely with Professor Norman Sadeh on LLMs and cybersecurity research, while also collaborating externally with Professor Ashique KhudaBukhsh (RIT) on exploring LLM political polarization and jailbreak-assisted toxic rabbit hole evaluations. My contributions to privacy-preserving machine learning and AI safety have been recognized through the AAAI Undergraduate Consortium Scholar and MITACS Research Scholar awards, further fueling my passion for bridging the gap between theoretical vulnerabilities and real-world implications.
scrollable: please scroll left or right →
Meta's Prompt-Guard, designed to protect LLMs against jailbreaks, vulnerable to an exploit with a 99.8% success...
The Register'Ignore previous instructions' thwarts Prompt-Guard model if you just add some good ol' ASCII code 32... But Priyanshu found that the fine-tuning...
News BytesAman Priyanshu, a bug hunter with enterprise AI application security firm Robust Intelligence, discovered this safety bypass...
Bank Info Security"The bypass involves inserting character-wise spaces between all English alphabet characters in a given prompt..."
ChannelE2E...the need for a multi-layer approach," said Robust Intelligence AI Security Researcher Aman Priyanshu.
Aman Priyanshu, Yash Maurya, Vy Tran, Suriya Ganesh Ayyamperumal
USENIX Conference on Privacy Engineering Practice and Respect
Nisha P Shetty, Balachandra Muniyal, Akshat Dokania, Sohom Datta, Manas Subramanyam Gandluri, Leander Melroy Maben, Aman Priyanshu
Security and Communication Networks
Nisha P Shetty, Balachandra Muniyal, Aman Priyanshu, Vedant Rishi Das
Journal of Cyber Security and Mobility
Aman Priyanshu, Supriti Vijay, Ayush Kumar, Rakshit Naidu, Fatemehsadat Mireshghallah
Pre-Print (In-Submission)
Aman Priyanshu, Supriti Vijay
Research & Reports Track at #ShowYourSkill, Coursera
Supriti Vijay, Aman Priyanshu
Updatable Machine Learning Workshop, ICML 2022
Aman Priyanshu, Sarthak Shastri, Sai Sravan Medicherla
43rd IEEE Symposium on Security and Privacy
Josy Elsa Varghese, Balachandra Muniyal, Aman Priyanshu
Computers & Electrical Engineering, Volume 98
Aman Priyanshu, Rakshit Naidu, Fatemehsadat Mireshghallah, Mohammad Malekzadeh
Privacy Preserving Machine Learning Workshop, ACM CCS 2021
Aman Priyanshu, Aleti Vardhan, Sudarshan Sivakumar, Supriti Vijay, Nipuna Chhabra
Workshop on Noisy User-generated Text (W-NUT), EMNLP 2021
Rakshit Naidu, Aman Priyanshu, Aadith Kumar, Sasikanth Kotti, Haofan Wang, Fatemehsadat Mireshghallah
Responsible Computer Vision Workshop, CVPR 2021 & Privacy Preserving Machine Learning Workshop, ACM CCS 2021
Aman Priyanshu, Mudit Sinha, Shreyans Mehta
3rd Workshop on Continual and Multimodal Learning for Internet of Things, IJCAI 2021
Aman Priyanshu, Rakshit Naidu
Machine Learning for Preventing and Combating Pandemics & Distributed and Private Machine Learning Workshops, ICLR 2021
Aman Priyanshu, Vedant Rishi Das, Shashank Rajiv Moghe, Harsh Rathod, Sai Sravan Medicherla, Mini Shail Chhabra, Sarthak Shastri
IEEE Sixth International Conference on Multimedia Big Data (BigMM)
scrollable: please scroll left or right →
We cracked a real-world differentially private synthetic data by linking public information to exposed PII overnight.
We built a keyword extractor, forgot about it, and somehow researchers are actually using it in their work.
I created a hierarchical topic modeling dataset from RedPajama, with 100k samples and 3 levels of topics.
I created a hierarchical topic modeling dataset from RedPajama, with 100k samples and 3 levels of topics.
I built a vanilla JavaScript library that lets you use AI APIs directly in browsers, no backend needed.
I mapped linguistic patterns in YC startup pitches like growing trees, and built a semantic search tool to explore them.
We broke AI safeguards by splitting harmful prompts into innocent sub-questions.
Feb 2023
May 2021 - Jun 2023
Feb 2021 - Dec 2022
Jan 2020 - Feb 2020
Key Courses: Prompt Engineering (17730), AI Governance (17716), Deep Learning (11785), Computer Technology Law (17562), Differential Privacy (17731), Information Security (17631), & Usability (17734)
Key Courses: Data Structures and Algorithms, Design and Analysis of Algorithms, Object Oriented Programming, Probability and Statistics, Computer Networks, Operating Systems, Database Management
June 2023
Specify your dataset of choice, and ProTaska-GPT will understand the dataset with tasks, tutorials, and actionable insights for it. Accelerate your data science journey with ease and efficiency! (Meant for people starting their journey into Data Science.)
October 2022
Built a python library, integrating semi-supervised attention for creating a few-shot & zero-shot domain adaptation technique for keyphrase extraction.
June 2022
Creating a Differential Privacy securing Synthetic Data Generation for tabular, relational and time series data.
May 2022
NERDA-Con is a python package, a pipeline for training Named Entity Recognition (NER) with Large Language Models bases by incorporating the concept of Elastic Weight Consolidation (EWC) into the NER fine-tuning NERDA pipeline.
August 2021
DP-HyperparamTuning offers an array of tools for fast and easy hypertuning of various hyperparameters for the DP-SGD algorithm. We proposed a novel, customizable reward function that allows users to define a single objective function for establishing their desired privacy-utility tradeoff.
August 2021
Created an unsupervised machine learning to extract contextually similar texts. The project was used in indexing Academic Literature, Law Precedents, and Financial Records. The project won Code Innovation Series - a Hackathon in association with GitHub.
July 2021
The app performs obstruction detection, spoof detection, blur detection and environment approval. Utilized Deep Neural Networks and Genetic Algorithms to achieve these goals in low computational time. The project won 1st place in HackRx 2.0 by Bajaj Finserv.
May 2021
DeCrise is an online platform that acts as an aggregator for public support/utility services which uses continual-federated-learning to create a quick response information retrieval system during a natural disaster. The project won 1st place in The ACM UCM Datathon.
April 2021
A social-media platform employing machine learning and differential privacy to promote civic engagement while protecting user-privacy. The project won under the Community & Civic Engagement for UC Berkeley's CalHacks Hackathon.
March 2024
Won the Spark Grant for our app that enhances speech for non-native English speakers, employing prompt-engineering function-calling (OpenAI GPT4/3.5) and Speech-to-Text (OpenAI Whisper), with features like audio-segmentation, speaker-recognition, and diarization.
September 2023
Won the Space-Themed track with our space trash collection project using Pareto optimization to balance time, fuel requirements, satellite movements, planetary alignment, and the trajectory of trash collectors for predicting monetary incentives.
February 2023
Selected as one of twelve individuals for the AAAI UC program, recognizing my research on Privacy and Fairness.
June 2022
Came second runners-up in #ShowYourSkill where we participated in the Research & Reports Track and creating a NLP augmented Machine Learning Application for women safety.
September 2021
Came runners-up in BobHacks where we built a pattern recognition API built on top of the MetaBob API. The API is able to assist users in tracking common errors and delivers pattern recognition on the MetaBob API.
August 2021
Innovation Series Hackathon was organized by Manipal Institute of Technology. Employed Document-Embedding for measuring contextual similarity between multiple pages and given search-queries.
July 2021
Used Deep Learning and Classical Image processing to achieve a face verification and profile-rank estimation task. The methodology out-performed classic Deep Learning methods.
May 2021
Built DeCrise, an online platform that acts as an aggregator for public support/utility services for fast-response during a major crisis or disaster.
April 2021
Won under the Community & Civic Engagement track. Built Voix, an anonymous platform for uplifting communities and promoting civic participation using privacy-enabled machine learning.
September 2020
Employed skip-connections to generate high-performance model for furniture identification in IECSE x VISION competition.
August 2020
Came runners-up in IEEE Grand-Challenge for harassment detection on tweets. Used Elementary Classifiers for Sentiment Analysis. The team was invited to present at IEEE BigMM conference.
January 2020
Selected as one of the recipients of the Intel Edge AI Scholarship Program. Learned about Machine Learning Implementation on the Edge.
scrollable: please scroll left or right →
Understand and play with federated learning hyperparams! In-browser tensorflow-js simulation of FedAvg to understand and gain intuition about IID and Non-IID Federated Learning settings.
A unique twist on classic Tetris where players manage a privacy budget to reveal blocks, demonstrating differential privacy concepts through gameplay. Experience privacy-utility tradeoffs in an engaging way.
An interactive game exploring machine learning unlearning and fairness concepts. Players select data points that least impact the dataset, providing hands-on experience with data removal and model fairness considerations.