Education

• Ph.D in Computer Science, Tulane University, 2023 (expected)
• M.S. in Computer Science, The Ohio State University, 2019
• B.Tech in Computer Engineering, Thapar University, 2015

Publications

• “Understanding the Limits of Unsupervised Domain Adaptation via Data Poisoning”. Code
  Akshay Mehra, Bhavya Kailkhura, Pin-Yu Chen and Jihun Hamm.
  Neural Information Processing Systems (NeurIPS) 2021.
• “How Robust are Randomized Smoothing based Defenses to Data Poisoning?”. Code
  Akshay Mehra, Bhavya Kailkhura, Pin-Yu Chen and Jihun Hamm.
  Computer Vision and Pattern Recognition (CVPR) 2021.
• “Penalty Method for Inversion-Free Deep Bilevel Optimization”. Code
  Akshay Mehra, Jihun Hamm.
  Asian Conference on Machine Learning (ACML) 2021.
• “Fast Interactive Image Retrieval using large-scale unlabeled data”.
  Akshay Mehra, Jihun Hamm and Mikhail Belkin.
  arXiv (2018).
• “Machine vs Machine: Minimax-Optimal Defense Against Adversarial Examples”.
  Jihun Hamm and Akshay Mehra.
  arXiv (2017).

Work experience

• Research Intern at Lawrence Livermore National Laboratory: Summer 2020
  • Supervisors: Dr. Bhavya Kailkhura
  • Data poisoning attacks against certified defenses.

    • We studied the problem of using data poisoning attacks to affect the robustness guarantees of classifiers trained using certified defense methods.

    • We proposed a bilevel-optimization based attack which can generate poison data against several robust training and certification methods. We specifically used the attack to highlight the vulnerability of randomized smoothing based certified defenses to data poisoning.

    • We demonstrated the effectiveness of our attack in reducing the certifiable robustness obtained using randomized smoothing on models trained with state-of-the-art certified defenses such as Gaussian data augmentation, SmoothAdv and MACER. For these training methods our attack reduced the average certified radius of the target class by more than 30\%.

• Summer Intern at Avata Intelligence: Summer 2019
  • Supervisors: Dr. Manish Jain & Dr. Matthew Brown
  • Train and Test time attacks on Neural Networks:

    • Evaluated effectiveness of various test time attacks on fully connected and convolutional neural networks. Compared the performance of these methods against a margin-based attack. Our margin-based attack produces adversarial examples with lesser distortion and is twice as fast when compared to the CW attack.

    • Compared performance of data poisoning attacks generated by solving Bilevel Optimization against other methods such as reverse mode automatic differentiation. Our method for solving Bilevel Optimization produces significantly better results using only a about 6% poisoning points.

• Data Scientist Intern at Microsoft: Summer 2017
  • Supervisors: Dr. Mohamed Abdel-Hady & Dr. Debraj GuhaThakurta
  • Entity Extraction from Bio-Medical Data, to identify Drugs, Diseases etc. from medical records. Published here.

    • Trained Word2Vec model using 15 million Medline Abstracts to extract word embeddings for medical entities using Spark in < 90 mins.

    • Build a LSTM based recurrent neural network to train the entity extractor using Keras with Tensorflow backend. RNN initialized with Medline embeddings outperformed the RNN initialized with generic word embeddings trained on Google News. Our model detected 7012 entities correctly (out of 9475) with a F1 score of 0.73 compared to 5274 entities with F1 score of 0.61 obtained with RNN initialized with generic word embeddings. Code

• Software Engineer at Microsoft: June 2015 - Aug 2016
  • Secure License Keys (SLK), a project involving generation and distribution of product keys for almost all Microsoft products.

    • Optimization of product key inventory using various machine learning techniques. Obtained mean absolute percentage error of 5.84 for the next peak quantity and 1.54 for next peak time. Improved the system by about 25%.

    • Utilized telemetry data of several internal applications for Anomaly Detection for proactive monitoring of systems.

• Software Engineer Intern at Microsoft: Summer 2014
  • Reporting on Cloud, a Hybrid BI Portal, aimed at being a one stop shop for all reporting needs of a business user.

    • Developed a SharePoint portal and integrated existing reports with Power View. Showed how effective visualization helps business understand support contract renewal by premier customers of Microsoft. Provided actionable insights for the PFE’s to ensure majority contract renewals, using Machine Learning models.

Skills

• Programming, Databases, Tools:
  • C C++, Python, Tensorflow, Keras, MATLAB, Shell Scripting, Azure, SQL Server, Oracle, MySQL.
• Competitive Programming Handles:
  • Codechef rihaan1991, Hackerrank akshaymehra, Google Code Jam (akshay1622).