CV

Education

• Ph.D in Computer Science, University of Illinois Chicago, 2020-2025 (expected)
• M.S. in Computer Science, Brown University, 2018-2020
• B.S. in Computer Science, B.S. in Mathematics, Minor in Physics, Pennsylvania State University, 2014-2018

Work experiences

• Summer 2023: Applied Scientist Internship @ Amazon
  • Built models to predict free-trial to subscription conversion rate on Amazon digital products including Music Unlimited, Prime, Kindle Unlimited, Video Channels, and Audible.
• Summer 2022: Applied Scientist Internship @ Amazon
  • Built sequential event models to identify bad-actor account based on their activities at Amazon, thereby preventing abusive content from entering the community.
• Summer 2018: Machine Learning Engineer Intership @ Accenture AI Lab
  • Developed an attention-based OCR model using TensorFlow that recognizes text and numbers from receipts, facilitating the determination of reimbursement for employees.

Research experiences

• Poisoning Attack on Reinforcement Learning @ University of Illinois Chicago
  • Introduced a novel data poisoning attack on offline-to-online reinforcement learning, revealing its vulnerabilities.
    • Promoted distribution shift by leveraging bi-level optimization techniques, resulting in more than 20% return drop during online fine-tuning while maintaining integrity during offline training.
    • Validated the effectiveness of the attack across four environments, highlighting its ability to operate with minimal budget without white-box access to the victim model.
  • Supervised by Prof. Xinhua Zhang.
  • Accepted by UAI(2024).
• Poisoning Attack on Generative Model @ University of Illinois Chicago
  • Introduced a poisoning attack on generative models in the context of replay based continual learning such as DGR.
    • Poison attacked the DGR model through input-aware backdoor attack; promoted catastrophic forgetting on the victim model, leading to an accuracy drop on past tasks to below 10%.
    • Proved the attack to be escapable from Neural Cleanse, the SOTA in defending backdoor attacks.
  • Supervised by Prof. Xinhua Zhang.
  • Published in ICML(2023).
• Recommendation System @ Brown University
  • Introduced an innovative solution to the cold-start problem by leveraging side information such as age, gender, etc.
    • Theoretically validated the empirical bounds of Cartesian Rademacher complexity on nuclear norm constrained models by reformulating the empirical risk minimization (ERM) for bivariate regression into a linear combination of quadratic and linear terms.
    • Designed the experiment using bivariate regression model.
  • Supervised by Prof. Eli Upfal.
• Pattern Recognition @ Pennsylvania State University
  • Predicting extreme weather from radar images.
    • Extract skeleton information from radar images in the NOAA dataset.
    • Utilized OpenCV to identify bow echoes from skeleton information.
    • Designed a time series model for predicting severe weather conditions using the bow-echo information.
  • Supervised by Prof. James Wang.
• Cryptography @ Purdue University
  • Analyzed the robustness of algorithms against parallel attacks by computing the space-time complexity of Argon2iA, Argon2iB, and greedy attacks.
  • Supervised by Prof. Jeremiah Blocki.
  • Results published in CRYPTO(2019).

Publications

Offline Reward Perturbation Boosts Distributional Shift in Online RL

Yu, Z.*, Kang, S.*, & Zhang, X. (*equal contribution). (2024, July). Offline Reward Perturbation Boosts Distributional Shift in Online RL. In The 40th Conference on Uncertainty in Artificial Intelligence.

Poisoning Generative Replay in Continual Learning to Promote Forgetting

Kang, S., Shi, Z., & Zhang, X. (2023, July). Poisoning generative replay in continual learning to promote forgetting. In International Conference on Machine Learning (pp. 15769-15785). PMLR.

Data-Independent Memory Hard Functions: New Attacks and Stronger Constructions

Blocki, J., Harsha, B., Kang, S., Lee, S., Xing, L., & Zhou, S. (2019, August). Data-independent memory hard functions: New attacks and stronger constructions. In Annual International Cryptology Conference (pp. 573-607). Cham: Springer International Publishing.

Services