Amazon - Virginia Tech Initiative for Efficient and Robust Machine Learning

Peng Gao and Ruoxi Jia

Project Goals and Objectives.  Machine learning (ML) models can expose the private information of training data when confronted with privacy attacks. Specifically, a malicious user with black-box access to an ML-as-a-service platform can reconstruct the training data or infer the membership information simply by querying the ML model. Despite the pressing need for effective defenses against privacy attacks with black-box access, existing approaches have mostly focused on increasing the robustness of models and none of the approaches focus on the continuous protection of already deployed ML models.

Project Results.  In this project, we designed and implemented a stream-based system for real-time privacy attack detection for ML models. To facilitate wide applicability and practicality, our system defends black-box ML models against a wide range of privacy attacks in a platform-agnostic fashion: our system only passively monitors model queries without requiring the cooperation of the model owner or the AI platform. Our evaluations against a wide range of realistic privacy attacks demonstrated that our defense approach can successfully mitigate the attack success rate for all attacks within a practical time frame.

Meetings and Other Outcomes.  For this project, we (Peng Gao and Ruoxi Jia) were honored to collaborate with our Amazon sponsor, Anoop Kumar. During the project time frame, we scheduled several meetings and discussed this project and other collaborations. The paper has already been accepted at RAID 2023, a premier, well-recognized security conference, as a joint publication. We will plan to continue our discussion with Anoop and collaborate on other projects on AI security and large language models.

Ismini Lourentzou

Project Goals and Objectives.  This project focuses on developing advanced conversational embodied agents that can adaptively request assistance from humans, perform multiple tasks simultaneously, and effectively capture shared knowledge. The primary technical objective is to create general- purpose embodied agents that can understand instructions, interact seamlessly with hu- mans, predict human beliefs, and successfully complete various tasks in physical environments.

Project Results.  The first research task focuses on developing an embodied agent that can interact and collaborate with humans when completing embodied tasks by asking questions for assistance and subsequently updating action plans based on feedback. To address the first research task, an Embodied Learning-By-Asking (ELBA) agent has been developed.  In the second research task, we explore the emergence of language in multi-agent populations. Our findings show performance gains in several downstream tasks when a model is pretrained on emergent language corpora and subsequently fine-tuned on natural language corpora. Similar to natural language, we observe that emergent language models are adept at utilizing positional information to reuse tokens while maintaining their descriptive power. The third research task involves designing an interaction model that incorporates physiological influences and behavioral norms in human-agent collaboration. The objective is to enhance social awareness and improve social human-agent interaction.

Meetings and Other Outcomes.  Throughout this project, close collaboration and interactions with Amazon researchers have been vital in achieving our research goals. We continue our work in this area and in task driven conversational agents through partnerships with Amazon Alexa. Our team is selected as one of the 10 teams competing at the Amazon Taskbot Challenge.  This work has been submitted for publication at a top-tier machine learning conference. The first author, Ph.D. student Makanjuola Ogunleye, is one of the eight nationwide recipients of the Cadence Black Students in Technology Scholarship.

Walid Saad

Project Goals and Objectives.  The goal of this project is to design green, efficient, and robust FL algorithms over resource-constrained devices and wireless systems (e.g., 5G). We thus aim to answer the following fundamental questions: 1) What are possible techniques for designing FL algorithms that can decrease the computational complexity of the training and inference? 2) What are the involved tradeoffs between learning accuracy and efficiency, model size, and convergence time? 3) How can we optimize those tradeoffs, and what is the optimal FL architecture and configuration for a given scenario? 4) How can we improve the robustness of FL algorithms for reliable operation on a number of IoT devices? This project answers these questions by bringing forward the new paradigm of communications, computing, and learning systems co-design.

Project Results.  Thus, far the project led to two key publications, one accepted book chapter, and one submitted Neurips paper.  In the book chapter, we primarily outlined the challenges, problems, and potential approaches of our project while focusing on how to design green FL systems that can operate in an energy-efficient manner. We then showed how such challenges require the co-design of computing, learning algorithms, and communication systems while also outlining some preliminary results using quantization for green FL. Based on this, in the Neurops submission, we developed a computation and communication-efficient FL framework SpaFL as shown in Fig. 1. To demonstrate the performance and efficiency of SpaFL, we conducted extensive experiments on three image classification datasets: FMNIST, CIFAR-10, and CIFAR-100 datasets. We compare SpaFL with six state-of-the-art baselines using convolutional neural network (CNN) models. In Table 1 and Fig. 2, we present the averaged accuracy, communication costs (Gbits), computation costs (the number of FLOPs), and convergence rate during training. From these results, we observe that SpaFL outperforms all baselines while using the least amount of communication costs and number of FLOPs. Hence, SpaFL is a resource-efficient FL framework and also robust to various data distributions across multiple devices. This algorithm can potentially admit many applications for deployment over resource-constrained devices like Alex Home Assistants.

Meetings and Other Outcomes.  During the course of the project, we have had a monthly meeting with Joe Wang from Amazon, who has helped us shape our contributions, while particularly guiding us on how to best evaluate and design our SpaFL algorithm. We expect those interactions to continue, and we will potentially aim at having joint publications.

The work has supported PhD student Minsu Kim who, along with Dr. Saad, attended the Amazon - VT workshop and interacted with several of the attendees. In particular, Minsu presented a poster of this work which was discussed with many Amazon attendees.

Dr. Saad has also discussed the outcomes of this project in several interviews on Fox and Friends, on the weather channel, as well as on an online platform called “Hello, Computer”.

Lifu Huang

The goal of this project is two-fold. First, it will develop an innovative, semi-parametric conversational framework that augments a large parametric conversation generation model with a large collection of information sources so desired knowledge is dynamically retrieved and integrated to the generative model, improving the adaptivity and scalability of the conversational agent toward open domain topics. Secondly, it will simulate fine-grained human judgments on machine-generated responses in multidimensions by leveraging instruction tuning on large-scale pre-trained models. The pseudo human judgments can be used to train a lightweight multidimensional conversation evaluator or provide feedback to conversation generation.

Ruoxi Jia

This project focuses on developing strategic data acquisition and pruning techniques that enhance training efficiency while addressing robustness against suboptimal data quality by creating targeted data acquisition strategies that optimize the collection of the most valuable and informative data for a specific task, designing data pruning methods to eliminate redundant and irrelevant data points, and assessing the impact of these approaches on computational costs, model performance, and robustness. When successfully completed it will optimize the data-for-AI pipeline by accelerating the development of accurate and responsible machine learning models across various applications.

Ming Jin 

Through the integration of reinforcement learning and game theory, the project aims to develop a fresh framework to ensure the safe and effective operation of interactive systems, such as conversational robots like ChatGPT. Importantly, the framework is designed to align with the needs and preferences of all relevant stakeholders, including users and service providers, each holding a vested interest in the system's performance.