Publications

LLM panels running the Delphi protocol achieve strong calibration (r=0.87–0.95) against benchmark ground truth and align closely with human expert panels, reducing elicitation time from months to minutes.

MIBP-Cert uses mixed-integer bilinear programming to compute sound, deterministic robustness bounds during training, handling complex threat models including discrete and continuous data perturbations.

First certification framework guaranteeing pixel-level robustness for any black-box attribution method via randomized smoothing, with new metrics for certified robustness, localization, and faithfulness.

FullCert is the first end-to-end certifier providing deterministic robustness bounds against both training-time poisoning attacks and inference-time adversarial examples jointly.

Adaptive hierarchical certification for semantic segmentation relaxes abstentions to coarser label levels, achieving higher certified information gain and lower abstain rates than flat certification.

Fallback strategies in certified neural networks can be deliberately triggered by backdoor attacks, causing up to 100% of inputs to be rejected — a novel availability threat.

3DCertify, the first verifier for point cloud models, certifies robustness against a wide range of semantic 3D transformations for both classification and part segmentation.

A GAN that synthesizes CNN features conditioned on class-level semantic information, enabling effective generalized zero-shot learning without labeled examples of unseen classes.