|
I’m currently working in the Privacy Team at the Trustworthy Technology Lab, Huawei Munich Research Center, where I focus on the privacy and safety of large language models (LLMs). My current research includes studying privacy leakage in LLMs, Unlearning of Sensitive information, text anonymization, and understanding LLMs through mechanistic interpretability. I graduated with a PhD in Computer Science in August 2022 from the Computer Vision Lab at EPFL. I was supervised by Dr. Mathieu Salzmann and Prof. Pascal Fua. My thesis focused on the robustness and interpretability of ML models. Following the completion of my PhD, I worked as a postdoctoral scientist at the Visual Intelligence for Transportation Lab (VITA) at EPFL, under the supervision of Prof. Alexandre Alahi, for eight months, until April 2023. Before joining EPFL in 2017, I spent two years at Samsung Research Bangalore working on mobile camera algorithms. Prior to that, I graduated from the Department of Electrical Engineering at IIT Kharagpur in 2015 with a dual degree (Master’s and Bachelor’s). During my undergraduate years, I interned at the University of Alberta, the University of Queensland, and Philips Research.
Email / CV / Google Scholar / Github / LinkedIn / Thesis / Thesis Slides |
|
|
My research interests is to learn models that are robust and interpretable, especially for safety and security-related applications. Currently, my research focuses on improving the privacy of Large Language Models (LLMs). During my PhD, I focused on understanding the weaknesses of deep neural networks, especially in handling unexpected or adversarial situations, to make them more robust. My research covered topics like explainable models, transfer-based black-box attacks, attack detection, adversarial defenses, anomaly detection, and testing disentangled representations. While at VITA, I worked on human pose estimation, tracking, and re-identification, mainly in the context of team sports analytics. |
|
Krishna Kanth Nakka Innosuisse VITA-Dartfish, Sep 2022 - April 2023 The objective of this project is to enhance sports player tracking through a unified framework. This involves detecting and tracking semantic keypoints and utilizing re-identification techniques to enhance long-term tracking, especially when players go out of view. |
|
Krishna Kanth Nakka, Mathieu Salzmann Preprint, 2022 Paper Our analyses show that disentanglement in the three state-of-the-art disentangled representation learning frameworks is far from complete, and that their pose codes contain significant appearance information |
|
Krishna Kanth Nakka, Mathieu Salzmann Paper Adversarial Robustness Workshop, European Conference on Computer Vision (ECCV), 2022 We propose to learn to generate a single perturbation from the object template only, that can be added to every search image and still successfully fool the tracker for the entire video. As a consequence, the resulting generator outputs perturbations that are quasi-independent of the template, thereby making them universal perturbations. |
|
Krishna Kanth Nakka, Mathieu Salzmann Neural Information and Processing Systems (NeurIPS), 2021 arXiv / code We show that generators trained with mid-level feature separation loss transfers significantly better in cross-model, cross-domain and cross-task setting |
|
Krishna Kanth Nakka, Mathieu Salzmann Asian Conference on Computer Vision (ACCV), 2020 arXiv / code / Slides We improve the robustness by introducing an attention-based regularization mechanism that maximally separates the latent features of discriminative regions of different classes while minimizing the contribution of the non-discriminative regions to the final class prediction. |
|
Krishna Kanth Nakka, Mathieu Salzmann European Conference on Computer Vision (ECCV), 2020 [Spotlight] arXiv / code / Slides We show that the resulting networks are sensitive not only to global attacks, where perturbations affect the entire input image, but also to indirect local attacks where perturbations are confined to a small image region that does not overlap with the area that we aim to fool. |
|
Krzysztof Lis, Krishna Kanth Nakka, Pascal Fua and Mathieu Salzmann International Conference on Computer Vision (ICCV) , 2019 arXiv / code / Poster We rely on the intuition that the network will produce spurious labels in regions depicting unexpected anomaly objects. Therefore, resynthesizing the image from the resulting semantic map will yield significant appearance differences with respect to the input image which we detect through an auxiliary network |
|
Krishna Kanth Nakka and Mathieu Salzmann Explainable and Interpretable AI workshop, ICCV, 2018 [Oral] arXiv / Slides We build upon the intuition that, while adversarial samples look very similar to real images, to produce incorrect predictions, they should activate codewords with a significantly different visual representation. We therefore cast the adversarial example detection problem as that of comparing the input image with the most highly activated visual codeword. |
|
Krishna Kanth Nakka and Mathieu Salzmann British Media Vision Conference (BMVC), 2018 arXiv / Poster we introduce an attentional structured representation learning framework that incorporates an image-specific attention mechanism within the aggregation process. |
|
SankarGanesh Jonna, Krishna Kanth Nakka and Rajiv Ranjan Sahay International Workshop on Video Segmentation, ECCV, 2016 [Oral] arXiv / Slides We use knowledge of spatial locations of fences to subsequently estimate occlusion-aware optical flow. We then fuse the occluded information from neighbouring frames by solving inverse problem of denoising |
|
SankarGanesh Jonna, Krishna Kanth Nakka and Rajiv Ranjan Sahay Journal of the Optical Society of America A (JOSA A) , 2016 arXiv / PDF Our approach of defencing is as follows: (i) detection of spatial locations of fences/occlusions in the frames of the video, (ii) estimation of relative motion between the observations, and (iii) data fusion to fill in occluded pixels in the reference image. We assume the de-fenced image as a Markov random field and obtain its maximum a posteriori estimate by solving the corresponding inverse problem. |
|
SankarGanesh Jonna, Krishna Kanth Nakka and Rajiv Ranjan Sahay Asian Conference on Pattern Recognition (ACPR), , 2015 arXiv / PDF / Poster We propose a semi-automated de-fencing algorithm using a video of the dynamic scene. The inverse problem offence removal is solved using split Bregman technique assuming total variation of the de-fenced image as the regularization constraint. |
|
Nilanjan Ray, Satarupa Mukherjee, Krishna Kanth Nakka, Scott T. Acton, Silvia S. Blanker Signal and Information Processing, GlobalSIP, 2014 arXiv / PDF We proposing a framework for user interactive segmentation of MRI of human leg muscles built upon the the strategy of bootstrapping with minimal supervision. |
|
Krishna Kanth Nakka Y. A. Tesiram, I. M. Brereton, M. Mobli and J. R. Harmer Physical Chemistry Chemical Physics (PCCP), 2014 Paper / PDF / Supp We show through non-linear sampling scheme with maximum entropy reconstruction technique in HYSCORE, the experimental times can be shortened by approximately an order of magnitude as compared to conventional linear sampling with negligible loss of information |
|
I'm deeply grateful for the generous scholarships I received throughout my academic journey. Some of these scholarships include:
|
|
I have peer-reviewed over 100 articles, including:
|
|
|
|
Credits: Webpage template from Jon Barron. |