Federated Learning (FL) is a privacy-preserving, collaborative machine learning technique where multiple clients train a shared model on their private datasets without sharing the data. While offering advantages, FL is susceptible to backdoor attacks, where attackers insert malicious model updates into the model aggregation process. Compromised models predict attacker-chosen targets when presented with specific attacker-defined inputs. Backdoor defences generally rely on anomaly detection techniques based on Differential Privacy (DP) or require legitimate clean test examples at the server. Anomaly detection-based defences can be defeated by stealth techniques and generally require inspection of client-submitted model updates. DP-based approaches tend to degrade the performance of the trained model due to excessive noise addition during training. Methods that require legitimate clean data on the server require strong assumptions about the task and may not be applicable in real-world settings. In this work, we view the question of backdoor attack robustness through the lens of loss function optimal points to build a defence that overcomes these limitations. We propose Mode Connectivity Based Federated Learning (MCFL), which leverages the recently discovered property of neural network loss surfaces, mode connectivity. We simulate backdoor attack scenarios using computer vision benchmark datasets, including CIFAR10, Fashion MNIST, MNIST, and Federated EMNIST. Our findings show that MCFL converges to high-quality models and effectively mitigates backdoor attacks relative to baseline defences from the literature without requiring inspection of client model updates or assuming clean data at the server.

Federated learning (FL) is a distributed, privacy-preserving learning paradigm where a joint model is trained on private data stored on client devices. Data owners (clients) train models locally and then submit them to an aggregation server for incorporation into the joint model. Malicious clients can apply training time attacks, e.g., backdoor attacks, by submitting maliciously trained models. Prior work has shown that Differential Privacy (DP) can provide certified robustness to backdoor attacks; however, there are limited studies regarding DP parameter selection as a function of the model architecture. In this work, we show empirically that larger models (i.e., with more parameters) require stronger DP parameter settings to mitigate backdoor attacks. Furthermore, we present a framework that alters the FL training algorithm to preserve certified accuracy round-by-round and show empirically that it is superior to a model trainer selecting DP parameters ahead of time before training begins and with incomplete information about the attacker. Although tools from DP are used in our proposed framework, it is focused on backdoor attack mitigation and does not provide privacy guarantees.

Current remote sensing platforms generate a vast amount of imagery but the best current methods to index and retrieve that data require expensive and difficult to procure labels. In this paper, we aim to address this problem by presenting a performant content based image retrieval (CBIR) system that is capable of indexing and retrieval using only unlabelled data. We investigate the use of self-supervised learning, a method for end-to-end learning of visual features from large datasets. In particular, we investigate the performance of four state-of-the-art self-supervised learning methods: variational autoencoders, bidirectional GANs, colourisation networks and DeepCluster, and evaluate the quality of the representations learned on remote sensing CBIR problems. Experiments on two very high resolution datasets show that the best of these methods, DeepCluster, is able to achieve near parity with supervised transfer learning despite not using any label information.