PhD Student · Computer Science · Dartmouth College

Prathamesh Devadiga

JASPR Lab · Advised by Prof. Shawn Shan

I work on the security and privacy of generative AI: what models reveal about their training data, whether that data can be verifiably removed, and how deployed systems fail under attack.

CV Scholar GitHub LinkedIn Email
Photo of Prathamesh Devadiga
Hanover, NH
$ cat research.txt

I build attacks and audits for generative models. My work is organized around three questions:

  1. 01

    Training Data & Memorization

    What do models retain from the data they were trained on?

    I build query-efficient, black-box attacks that extract memorized training data from instruction-tuned models and production APIs, and study why prompt-level safety defenses fail to stop them.

  2. 02

    Provenance & Auditing

    Can we verify what a model learned, or unlearned?

    Machine unlearning often doesn't remove what it claims. I show that concept-erased diffusion models leave traces: comparing the original and edited checkpoints reveals which artist style was deleted.

    Deletion Scars · ECCV 2026 U&ME Workshop
  3. 03

    Security & Privacy of Deployed Models

    How do real-world systems fail under attack?

    I have worked on multi-agent defenses against jailbreaks and prompt injection, evaluation blind spots in legal LLMs, and adversarially robust malware detection.

    ICML 2026 AI4Law · GUARDIAN · KASPER (Applied Soft Computing)

About

I am a first-year PhD student in Computer Science at Dartmouth College, where I am a member of the JASPR Lab (Joint AI Security Privacy Research).

Before Dartmouth, I received my B.Tech in Computer Science from PES University, Bangalore. Along the way I was a research intern at Lossfunk (low-resource language modeling), an undergraduate researcher at IIT Indore (adversarially robust malware detection), and a Google Summer of Code 2025 contributor with UC Santa Cruz OSPO. I also founded Ādhāra AI Labs, an independent group working on efficient models and ML systems.

I am happy to talk about research and collaborations. The best way to reach me is by email.

News View all →

  • Started my CS PhD at Dartmouth College, joining the JASPR Lab with Prof. Shawn Shan.
  • Paper accepted at ECCV 2026 U&ME Workshop: Deletion Scars: Inferring Deleted Artist Styles from Paired Diffusion Models.
  • Paper accepted at ICML 2026 AI4Law Workshop: Resistant to Lawyers, Defeated by Disagreement.
  • Paper at the EACL 2026 LoResLM Workshop: Making Large Language Models Speak Tulu: Structured Prompting for an Extremely Low-Resource Language.
  • Joined Lossfunk as an AI Research Intern, working on language modeling for extremely low-resource languages.
  • Completed Google Summer of Code 2025 with UCSC-OSPO on billion-scale ANN embedding benchmarks.

Publications View all →

  1. ECCV
    Deletion Scars: Inferring Deleted Artist Styles from Paired Diffusion Models
    Prathamesh Devadiga
    In ECCV 2026 Workshop on Unlearning and Model Editing (U&ME), 2026
    Bib HTML
  2. ICML
    Resistant to Lawyers, Defeated by Disagreement: Evaluation Blindspots in Legal Language Models
    Prathamesh Devadiga and Advika Lakshman
    In ICML 2026 Workshop on AI for Law (AI4Law), 2026
    Bib
  3. NeurIPS
    SLMs as Compiler Experts: Auto-Parallelization for Heterogeneous Systems
    Prathamesh Devadiga and others
    In NeurIPS 2025 Workshop on Machine Learning for Systems, 2025
    Bib
  4. EACL
    Making Large Language Models Speak Tulu: Structured Prompting for an Extremely Low-Resource Language
    Prathamesh Devadiga and others
    In EACL 2026 Workshop on Language Models for Low-Resource Languages (LoResLM), 2026
    Bib

Writing View all →