Ambar Pal

About Me

I obtained my PhD from the Computer Science Department at the Johns Hopkins University, where my advisors were René Vidal and Jeremias Sulam. I was affiliated with the Mathematical Institute for Data Science. Previously I obtained my Masters degree from JHU, and my Bachelors degree from from IIIT Delhi, both in Computer Science. My CV can be found here.

I was an Amazon Fellow for 2023-24. I received the JHU MINDS Data Science Fellowship 3 times, for the years 2022, 2021 and 2019.

Research Interests

My research focusses on the theory and practice of Robustness in Machine Learning. My central philosophy is that incorporating structural constraints from data can efficiently mitigate vulnerabilities in current ML systems to malicious agents. On the theoretical side, I have built frameworks formalizing how exploiting structure in data (e.g., low-dimensionality) can lead to improved formal guarantees for robustness. On the practical side, I have demonstrated the utility of such data-driven constraints for systems for image classification, graph classification, and facial recognition.

Publications and Preprints

Provable Robustness

• Certified Robustness against Sparse Adversarial Perturbations via Data Localization [Preprint]
  Ambar Pal, René Vidal, Jeremias Sulam


• Adversarial Examples Might be Avoidable: The Role of Data Concentration in Adversarial Robustness [NeurIPS 2023]
  Ambar Pal, Jeremias Sulam, René Vidal
• Poster in [DeepMath 2023]


• Understanding Noise-Augmented Training for Randomized Smoothing [TMLR 2023]
  Ambar Pal, Jeremias Sulam
• Shorter version in [DeepMath 2022], [AI-X Symposium 2023]


• Certified Defenses Against Near-Subspace Unrestricted Adversarial Attacks [ECCVW 2022]
  Ambar Pal, René Vidal.
• Poster in [SLowDNN 2023]


• A Game Theoretic Analysis of Additive Adversarial Attacks and Defenses [NeurIPS 2020]
  Ambar Pal, René Vidal

Theory of Deep Learning

• On the Regularization Properties of Structured Dropout [CVPR 2020]
  Ambar Pal, Connor Lane, René Vidal, Benjamin D. Haeffele

Empirical Robustness

• Identifying Physically Realizable Triggers for Backdoored Face Recognition Networks [ICIP 2021]
  Ankita Raj, Ambar Pal, Chetan Arora


• Principled Attacks to Graph Neural Networks
  Ambar Pal, Julio Hurtado, Marcel Nassar, Nesreen K. Ahmed, René Vidal


• Making Deep Neural Network Fooling Practical [ICIP 2018]
  Ambar Pal, Chetan Arora

Object Detection, Visual Question Answering, and Others

• On Utilizing Relationships for Transferrable, Few-Shot Object Detection [ACVC 2022]
  Ambar Pal, Arnau Ramisa, Amit Kumar K C, René Vidal


• An Empirical Evaluation of Visual Question Answering for Novel Objects [CVPR 2017]
  Santhosh Kumar R, Ambar Pal, Gaurav Sharma, Anurag Mittal


• Exploiting independent visual and textual data sources to improve multi-modal methods for description and querying of visual data [IIITD Thesis 2018]
  Ambar Pal, Gaurav Sharma, Chetan Arora


• The Internet of Things: Perspectives on Security from RFID and WSN [Preprint]
  Ayush Shah, H. B. Acharya, Ambar Pal.

News

Past Life

In the past, I have been fortunate to have worked with several wonderful people on several exciting problems. Chetan Arora[IIT Delhi] was my undergraduate advisor, and we analysed the robustness of image classification and facial recognition models to various types of physically realizable adversarial perturbations. I interned with Yan Liu[USC] on the problem of survival prediction for cancer patients using machine learning. I worked with Gaurav Sharma at IIT Kanpur, where together with Santhosh we explored the limits of visual question answering. I have worked with H.B. Acharya[RIT] on classifying and identifying security aspects of the Internet of Things. With Somitra Sanadhya[IIT Ropar], I have explored and implemented differential attacks on the cryptographic hash function SHA-2. Finally, I have worked with Rajiv Raman[IIIT Delhi] on the planar support problem in discrete geometry.