Siddhartha Datta

I am pursuing a PhD in Computer Science at the University of Oxford under the supervision of Prof Sir Nigel Shadbolt. My thesis is focused on the adaptation of deep neural networks against distribution shifts. I study this theoretically through the construction of compressed low-loss subspaces to handle joint shifts manifesting label (adversarial/backdoor attacks), domain shift, and task shift. I also study its applications in human-AI collaborative settings.


Publications

  • Siddhartha Datta, Nigel Shadbolt, (Acknowledgements: Jamie Hayes), Multiple Modes for Continual Learning, NeurIPS 2022 Workshop.
    PDF


  • Siddhartha Datta, Nigel Shadbolt, Interpolating Compressed Parameter Subspaces, NeurIPS 2022 Workshop.
    PDF


  • Siddhartha Datta, Konrad Kollnig, Nigel Shadbolt, GreaseVision: Rewriting the Rules of the Interface, NAACL 2022 Workshop.
    PDF / talk


  • Siddhartha Datta, Learn2Weight: Parameter Adaptation against Similar-domain Adversarial Attacks, COLING 2022.
    PDF / talk


  • Siddhartha Datta, Nigel Shadbolt, Low-Loss Subspace Compression for Clean Gains against Multi-Agent Backdoor Attacks, CVPR 2022 Workshop.
    PDF


  • Siddhartha Datta, Nigel Shadbolt, Backdoors Stuck At The Frontdoor: Multi-Agent Backdoor Attacks That Backfire, ICLR 2022 Workshop.
    PDF


  • Siddhartha Datta, Konrad Kollnig, Nigel Shadbolt, Mind-proofing Your Phone: Navigating the Digital Minefield with GreaseTerminator, ACM IUI 2022.
    PDF / video / talk


  • Siddhartha Datta, Nigel Shadbolt, Hiding Behind Backdoors: Self-Obfuscation Against Generative Models, Technical Report.
    PDF


  • Siddhartha Datta, Giulio Lovisotto, Ivan Martinovic, Nigel Shadbolt, Widen the Backdoor To Let More Attackers In, Technical Report.
    PDF


  • Konrad Kollnig, Siddhartha Datta, Nigel Shadbolt, Imagining, Studying and Realising A Less Harmful App Ecosystem, OII Connected Life 2022.
    PDF


  • Konrad Kollnig*, Siddhartha Datta*, Max Van Kleek [*equal contribution], I Want My App That Way: Reclaiming Sovereignty Over Personal Devices, ACM CHI LBW 2021.
    PDF / talk


  • Siddhartha Datta, DeepObfusCode: Source Code Obfuscation Through Sequence-to-Sequence Networks, AISC 2021.
    PDF / talk


  • Highlighted Projects

    • MemoryPalace

      Winning submisson at CalHacks 5.0 (Epson sponsor award). Our team developed a pair of display goggles that helps patients with Alzheimer's or other visual/mental disabilities in locating their personal belongings. Users verbalize the item they are searching for (typically items from the COCO image dataset), and all objects detected would be highlighted.

      code / video


    • BreakupBot

      BreakupBot is a therapeutic chatbot developed to help get over romantic breakups. Organically acquired 200+ users of varying demographics. The system, initially built to help people who could not get over breakups, uses scraped content from dating sites and builds a philosophical counselling knowledge graph as the decision tree for generating responses to user input.

      code / video


    • GIFShop Wizard

      Computer vision has been left out of the hands of many photoshopping enthusiasts and chatbot users alike. We brought automated GIF-editing functionality to the masses with GIFShop Wizard, including First Order of Motion, Foreground Removal, Fast/Segmented Style Transfer, and CycleGAN through a chatbot interface.

      code / video


    • GreaseDroid

      GreaseDroid makes it easier for Android users to make their apps fit their special needs by decoding and inserting modular patches.

      code


    • GreaseTerminator

      GreaseTerminator is an active, non-invasive intervention tool to patch dark patterns in operating systems and app ecosystems through interface re-rendering.

      code


    • FaceMap

      Interpolating/transferring real-time user faces onto semantically-segmented StreetView buildings/roads/sky/... while users navigate around the world seen as them.

      code


    • Music Generation through Food Detection, Assignment of Music Notes to Distinct Items, and Utensil Impact Detection

      An an augmented reality system/game where people can see highlights (bounding boxes) over their food, and tapping their food will produce a certain instrumental sound to synthesize music (including guitar chords, piano on different scales, and drums). Deployed on mixed reality lens with object detection.

      code / video


    • Location Network Fingerprinting: Network Analysis of Franchise Expansion

      After aggregating expansion data from sucessful franchises/chains from many industries (including fast food, supermarkets, coffee, ...), we developed network analysis algorithms that identified statistically-likely patterns or fingerprints of expansion success. Specifically, we found that certain succesful firms tended to expand in certain ways that consolidated their success. We found the results to be 98% statistically significant, and pitched the system as a service for firms intending to expand within the US but are strategically unsure how.

      code


    • HAIV: Computational Precision Medicine against HIV Drug Resistance

      A mutation prediction algorithm built to map out subsequent mutations of a strain of HIV virus in a person's bloodstream over n timesteps, and can be further used in optimizing antiretroviral treatment (allocating the right medication at the right time to avoid drug resistance).

      white paper


    • (UG/PG) Tutor, Computer Security, Advanced Security, Artificial Intelligence, Networks, Software Engineering, Design Patterns, 2020-
      Competitive Programming, 10+ Hackathon Winner (incl. CalHacks, AngelHacks, ETHGlobal), 2017-2020
      -->