Tribhuvanesh Orekondy

I am a machine learning researcher at Qualcomm AI research in Switzerland. My research interests are data-driven differentiable simulations, neural rendering, generative models, and black-box optimization.

I completed by PhD in 2020 at the Max Planck Institute for Informatics in Computer Vision and Machine Learning and was advised by Mario Fritz and Bernt Schiele. Previously, I graduated with a Master's degree in Computer Science from ETH Zürich.

Email  ·  Google Scholar  ·  Github  ·  LinkedIn  ·  MPI

News

Research

I'm broadly interested in Computer Vision and Machine Learning. Currently, I work on Neural Simulations and investigate data-driven and differentiable techniques for simulations (e.g., for physics, engineering). My current research work spans techniques in neural rendering, generative modelling, inverse problems, reinforcement learning, and black-box optimization. Previously, during my PhD, I focused on topics in trustworthy and reliable ML (adversarial ML, privacy-preserving techniques).
Probabilistic and Differentiable Wireless Simulation with Geometric Transformers
Probabilistic and Differentiable Wireless Simulation with Geometric Transformers
Thomas Hehn, Markus Peschl, Tribhuvanesh Orekondy, Arash Behboodi Johann Brehmer
arXiv, 2024
paper  ·  bibtex

Geometric DL transformer based surrogate simulator for modelling joint distributions between 3d environment and its influence on EM waves.

Simulating, Fast and Slow
Simulating, Fast and Slow: Learning Policies for Black-Box Optimization
Fabio Valerio Massoli, Tim Bakker, Thomas Hehn, Tribhuvanesh Orekondy, Arash Behboodi
arXiv, 2024
paper  ·  bibtex

Solving black-box optimization problems on stochastic simulators using neural surrogates and reinforcement learning.

Active Learning Policies
Active Learning Policies for Solving Inverse Problems
Tim Bakker, Thomas Hehn, Tribhuvanesh Orekondy, Arash Behboodi, Fabio Valerio Massoli
NeurIPS ReALML Workshop, 2023
paper  ·  bibtex

We reinforcement learn active learning policies to guide local-surrogate-based inverse problem optimisation.

Switching Policies
Switching policies for solving inverse problems
Tim Bakker, Fabio Valerio Massoli, Thomas Hehn, Tribhuvanesh Orekondy, Arash Behboodi
NeurIPS Deep Inverse Workshop, 2023
paper  ·  bibtex

Reinforcement learning policies to guide surrogate-based inverse problem optimisation.

Transformer Path Loss
Transformer-Based Neural Surrogate for Link-Level Path Loss Prediction from Variable-Sized Maps
Thomas Hehn, Tribhuvanesh Orekondy, Ori Shental, Arash Behboodi Juan Bucheli, Akash Doshi, June Namgoong, Taesang Yoo, Ashwin Sampth, Joseph Soriaga,
GLOBECOM, 2023
paper  ·  bibtex

Transformer-based neural surrogate to model mmWave path losses in dense urban scenarios.

WiNeRT
WiNeRT: Towards Neural Ray Tracing for Wireless Channel Modelling and Differentiable Simulations
Tribhuvanesh Orekondy, Prateek Kumar, Shreya Kadambi, Hao Ye, Joseph Soriaga, Arash Behboodi
ICLR, 2023
paper  ·  talk  ·  dataset  ·  bibtex

NeRF-like neural surrogate to model differentiable wireless electromagnetic propagation effects in indoor environments.

MIMO-GAN
MIMO-GAN: Generative MIMO Channel Modeling
Tribhuvanesh Orekondy, Arash Behboodi, Joseph Soriaga
ICC, 2022
paper  ·  bibtex

GAN-based surrogate simulator to learn distributions of channel impulse response complex waveforms.

InfoScrub
InfoScrub: Towards Attribute Privacy by Targeted Obfuscation
Hui-Po Wang, Tribhuvanesh Orekondy, Mario Fritz
CVPR (Fair, Trusted, and Data Efficient Computer Vision workshop), 2021
paper  ·  bibtex

An image obfuscation network to remove privacy attribute information (such as by inverting, or maximizing uncertainty), while retaining image fidelity.

Membership Inference

Differential Privacy Defenses and Sampling Attacks for Membership Inference
Shadi Rahimian, Tribhuvanesh Orekondy, Mario Fritz
AISec, 2021
paper  ·  bibtex

Differential Privacy approaches to defend against membership inference attacks.

InfoScrub
GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generators
Dingfan Chen, Tribhuvanesh Orekondy, Mario Fritz
NeurIPS, 2020
paper  ·  bibtex

A novel GAN to allow releasing sanitized forms of data with rigorous differential privacy guarantees.

Prediction Poisoning
Prediction Poisoning: Towards Defenses Against DNN Model Stealing Attacks
Tribhuvanesh Orekondy, Bernt Schiele Mario Fritz
ICLR, 2020
paper  ·  project page  ·  bibtex

An optimization-based defense against model stealing attacks, with perturbations crafted to poison resulting gradient signals.

Knockoff
Knockoff Nets: Stealing Functionality of Black-Box Models
Tribhuvanesh Orekondy, Bernt Schiele Mario Fritz
CVPR, 2019
paper  ·  poster  ·  extended abstract (CV-COPS@CVPR)  ·  project page  ·  bibtex

Vision models encode meaningful information in predictions even on out-of-distribution natural images. We exploit this property to steal functionality of complex vision models.

Deanonymization

Gradient-Leaks: Understanding and Controlling Deanonymization in Federated Learning
Tribhuvanesh Orekondy, Seong Joon Oh, Yang Zhang, Bernt Schiele Mario Fritz
FL NeurIPS, 2019
paper  ·  poster  ·  talk  ·  bibtex

Gradient parameter deltas in Federated Learning encodes user bias statistics of participating devices, raising deanonymization concerns.

Visual Redactions

Connecting Pixels to Privacy and Utility: Automatic Redaction of Private Information in Images
Tribhuvanesh Orekondy, Mario Fritz, Bernt Schiele
CVPR, 2018 (Spotlight)
paper  ·  poster  ·  project page  ·  video  ·  bibtex

Automatic method to identify and redact a broad range of private information spanning multiple modalities in visual content.

VPA

Towards a Visual Privacy Advisor: Understanding and Predicting Privacy Risks in Images
Tribhuvanesh Orekondy, Bernt Schiele, Mario Fritz
ICCV, 2017
paper  ·  poster  ·  extended abstract (VSM@ICCV)  ·  project page  ·  bibtex

An approach to understand and predict a wide spectrum of privacy risks in images.

Hades

HADES: Hierarchical Approximate Decoding for Structured Prediction
Tribhuvanesh Orekondy (under supervision of Martin Jaggi, Aurelien Lucchi, Thomas Hoffman )
Master Thesis, 2016
paper  ·  project page  ·  bibtex

A fast structured output learning algorithm, which works by approximately decoding oracles to various extents.

Academic Activities

  • Reviewing: CVPR '19, CV-COPS '19, TPAMI '19, ICCV '20, AAAI '20, CVPR '20, ECCV '20, NeurIPS '20, IJCV '20, WACV '21, CVPR '21, ICLR '21 (Outstanding reviewer award)
  • Teaching Assistant: Machine Learning in Cyber Security, 2018, 2019
  • Thesis co-supervision: Shadi Rahimian (MSc., University of Saarland), Jonas Klesen (BSc., University of Saarland)