Peer-to-Peer federated learning is a distributed machine learning paradigm with a primary goal of learning a well-performing global model by collaboratively learning a shared model at different data hubs without the need of sharing data. Due to its immense practical applications, there is growing attention towards various challenges of efficient federated learning including communication efficiency, assumptions on connectivity, data heterogeneity, enhanced privacy, etc. In this paper, we address the problem of dynamic network topologies in federated learning. We present a technique to help new participants in Peer-to-Peer federated learning reach best possible accuracy by leveraging learning at other devices in a communication efficient manner. We model the costs in federated learning and apply a graph theoretical framework to show that one can draw from a range of graph-based algorithms to construct an efficient communication algorithm on a connected network, thereby matching the inference efficiency of centralized federated learning. We conduct experiments with varied graph formations and sizes to validate our claims.
Deep Neural Networks (DNNs) have emerged as a prominent set of algorithms for complex real-world applications. However, state-of-the-art DNNs require a significant amount of data and computational resources to train and generalize well for real-world scenarios. This dependence of DNN training on a large amount of computational and memory resources has increased the use of Machine Learning as a Service (MLaaS) or third-party resources for training large models for complex applications. Specifically, the drift of the deep learning community towards self-supervised learning for learning better representations directly from large amounts of unlabeled data has amplified the computational and memory requirements for machine learning. On the one hand, the availability of MLaaS (or third-party resources) alleviates this issue. On the other hand, it opens up avenues for a new set of vulnerabilities, where an adversary (someone from a third party) can infect the model with malicious functionality that is triggered only with specific input patterns. Such attacks are usually referred to as Trojan or backdoor attacks and are very stealthy and hard to detect. In this paper, we highlight the complete attack surface that can be exploited to inject hidden malicious functionality (backdoors) in machine learning models. We classify the attacks into two major categories, i.e., poisoning attacks and non-poisoning attacks, and present state-of-the-art works related to each. Towards the end of the article, we highlight the limitations of existing techniques and cover some of the key challenges in developing stealthy and robust real-world backdoor attacks.
Adversarial patch-based attacks have shown to be a major deterrent towards the reliable use of machine learning models. These attacks involve the strategic modification of localized patches or specific image areas to deceive trained machine learning models. In this paper, we propose DefensiveDR, a practical mechanism using a dimensionality reduction technique to thwart such patch-based attacks. Our method involves projecting the sample images onto a lower-dimensional space while retaining essential information or variability for effective machine learning tasks. We perform this using two techniques, Singular Value Decomposition and t-Distributed Stochastic Neighbor Embedding. We experimentally tune the variability to be preserved for optimal performance as a hyper-parameter. This dimension reduction substantially mitigates adversarial perturbations, thereby enhancing the robustness of the given machine learning model. Our defense is model-agnostic and operates without assumptions about access to model decisions or model architectures, making it effective in both black-box and white-box settings. Furthermore, it maintains accuracy across various models and remains robust against several unseen patch-based attacks. The proposed defensive approach improves the accuracy from 38.8% (without defense) to 66.2% (with defense) when performing LaVAN and GoogleAp attacks, which supersedes that of the prominent state-of-the-art like LGS [19] (53.86%) and Jujutsu [7] (60%).
Adversarial attacks have emerged as a major challenge to the trustworthy deployment of machine learning models, particularly in computer vision applications. These attacks vary in their potency and can be implemented through both white-box and black-box approaches. Practical attacks include methods to manipulate the physical world and enforce adversarial behavior in the targeted neural network models. Multiple different approaches to mitigate these attacks have been proposed in the literature, each with distinct strengths, limitations, and trade-offs. In this survey, we present a comprehensive systematization of knowledge on adversarial defenses, focusing on two key computer vision tasks: image classification and object detection. We review and categorize state-of-the-art adversarial defense techniques to facilitate clearer comparative insights. Additionally, we provide a schematic representation of these categories within the context of the overall machine learning pipeline, facilitating clearer understanding and benchmarking of defenses. Furthermore, we map these defenses to specific adversarial attacks and datasets where their effectiveness has been demonstrated, offering practical insights for researchers and practitioners. This study is necessary for understanding the extent to which available defenses can address the adversarial threats, their inherent limitations, and open challenges, thereby guiding future research toward building trustworthy AI systems suitable for real-world deployment.