AI safety & model auditing

My efforts in the area of AI SEA (Safety, Ethics and Alignment) have broadly been in these two areas:
Red-Teaming + Adversarial attacks and Dataset auditing + bias mitigation and Model auditing.

A sampling of my research publications spanning these two areas : Model auditing, Red-Teaming and Adversarial attacks

1. Model auditing, Red-Teaming and Adversarial attacks

   1. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models

   2. Auditing saliency cropping algorithms. Project page [Video]

   3. Model weight theft with just noise inputs: The curious case of the petulant attacker

   4. Understanding adversarial robustness through loss landscape geometries

   5. Vulnerability of deep learning-based gait biometric recognition to adversarial perturbations

   6. On Lyapunov exponents and adversarial perturbation

   7. On detecting adversarial inputs with entropy of saliency maps

   8. On grey-box adversarial attacks and transfer learning

   9. Art-attack! on style transfers with textures, label categories and adversarial examples

   10. SIMUni: Sampling Impostors from Misfit Universal Background Models in accelerometric gait biometric verification

   11. Smile in the face of adversity much? A print based spoofing attack

   12. OODles of ODDs: The landscape of Out-of-distribution vulnerabilities of vision models

   13. Did They Direct the Violence or Admonish It? A Cautionary Tale on Contronomy, Androcentrism and Back-Translation Foibles [Video]

2. Dataset curation, auditing and bias mitigation

   1. Multimodal datasets: misogyny, pornography, and malignant stereotypes

   2. Large image datasets: A pyrrhic win for computer vision?

   3. Into the laion’s den: Investigating hate in multimodal datasets

   4. On Hate Scaling Laws For Data-Swamps

   5. The blood diamond effect in neural art: On ethically troublesome images of the imagenet dataset [Video]

   6. Covering up bias in CelebA-like datasets with Markov blankets: A post-hoc cure for attribute prior avoidance

   7. JFT-300M: The mystery box

PS: My more recent AI safety chronicles can be found here

JFT 300M

Translation foibles

Saliency Image Cropping audits

Nueral art and blood diamond effect