Search

Federated learning allows multiple users to collaboratively train a shared classification model while preserving data privacy. This approach, where model updates are aggregated by a central server, was shown to be vulnerable to poisoning backdoor attacks: a malicious user can alter the shared model to arbitrarily classify specific inputs from a given class. In this article, we analyze the effects of backdoor attacks on federated meta-learning, where users train a model that can be adapted to different sets of output classes using only a few examples. While the ability to adapt could, in principle, make federated learning frameworks more robust to backdoor attacks (when new training examples are benign), we find that even one-shot attacks can be very successful and persist after additional training. To address these vulnerabilities, we propose a defense mechanism inspired by matching networks, where the class of an input is predicted from the similarity of its features with a support set of labeled examples. By removing the decision logic from the model shared with the federation, the success and persistence of backdoor attacks are greatly reduced. [ABSTRACT FROM AUTHOR]

Various sensor network measurement studies have reported instances of transient faults in sensor readings. In this work, we seek to answer a simple question: How often are such faults observed in real deployments? We focus on three types of transient faults, caused by faulty sensor readings that appear abnormal. To understand the prevalence of such faults, we first explore and characterize four qualitatively different classes of fault detection methods. Rule-based methods leverage domain knowledge to develop heuristic rules for detecting and identifying faults. Estimation methods predict "normal" sensor behavior by leveraging sensor correlations, flagging anomalous sensor readings as faults. Time-series-analysis-based methods start with an a priori model for sensor readings. A sensor measurement is compared against its predicted value computed using time series forecasting to determine if it is faulty. Learning-based methods infer a model for the "normal" sensor readings using training data, and then statistically detect and identify classes of faults. We find that these four classes of methods sit at different points on the accuracy/robustness spectrum. Rule-based methods can be highly accurate, but their accuracy depends critically on the choice of parameters. Learning methods can be cumbersome to train, but can accurately detect and classify faults. Estimation methods are accurate, but cannot classify faults. Timeseries-analysis-based methods are more effective for detecting short duration faults than long duration ones, and incur more false positives than the other methods. We apply these techniques to four real-world sensor datasets and find that the prevalence of faults as well as their type varies with datasets. All four methods are qualitatively consistent in identifying sensor faults, lending credence to our observations. Our work is a first step towards automated online fault detection and classification. [ABSTRACT FROM AUTHOR]

We study an optimization problem that arises in the context of data placement in a multimedia storage system. We are given a collection of M multimedia objects (data objects) that need to be assigned to a storage system consisting of N disks d1, d2… , dN .We are also given sets U1,U2, …,UM such that Ui is the set of clients seeking the ith data object. Each disk dj is characterized by two parameters, namely, its storage capacity Cj which indicates the maximum number of data objects that may be assigned to it, and a load capacity L j which indicates the maximum number of clients that it can serve. The goal is to find a placement of data objects to disks and an assignment of clients to disks so as to maximize the total number of clients served, subject to the capacity constraints of the storage system. We study this data placement problem for two natural classes of storage systems, namely, homogeneous and uniform ratio. We show that an algorithm developed by Shachnai and Tamir [2000a] for data placement achieves the best possible absolute bound regarding the number of clients that can always be satisfied. We also show how to implement the algorithm so that it has a running time of O((N + M) log(N + M)). In addition, we design a polynomial-time approximation scheme, solving an open problem posed in the same paper. [ABSTRACT FROM AUTHOR]