FREPD: A Robust Federated Learning Framework on Variational Autoencoder

Federated learning is an ideal solution to the limitation of not preser-ving the users’privacy information in edge computing.In federated learning,the cloud aggregates local model updates from the devices to generate a global model.To protect devices’privacy,the cloud is designed to have no visibility into how these updates are generated,making detecting and defending malicious model updates a challenging task.Unlike existing works that struggle to tolerate adversarial attacks,the paper manages to exclude malicious updates from the glo-bal model’s aggregation.This paper focuses on Byzantine attack and backdoor attack in the federated learning setting.We propose a federated learning frame-work,which we call Federated Reconstruction Error Probability Distribution(FREPD).FREPD uses a VAE model to compute updates’reconstruction errors.Updates with higher reconstruction errors than the average reconstruction error are deemed as malicious updates and removed.Meanwhile,we apply the Kolmogorov-Smirnov test to choose a proper probability distribution function and tune its parameters to fit the distribution of reconstruction errors from observed benign updates.We then use the distribution function to estimate the probability that an unseen reconstruction error belongs to the benign reconstruction error distribution.Based on the probability,we classify the model updates as benign or malicious.Only benign updates are used to aggregate the global model.FREPD is tested with extensive experiments on independent and identically distributed(IID)and non-IID federated benchmarks,showing a competitive performance over existing aggregation methods under Byzantine attack and backdoor attack.

Optimizing partitioning strategies for faster inverted index compression

The inverted index is a key component for search engines to manage billions of documents and quickly respond to users' queries. Whereas substantial effort has been devoted to reducing space occupancy and decoding speed, the encoding speed when constructing the index has been overlooked. Partitioning the index aligning to its clustered distribution can effectively minimize the compressed size while accelerating its construction procedure. In this study, we introduce compression speed as one criterion to evaluate compression techniques, and thoroughly analyze the performance of different partitioning strategies. Optimizations are also proposed to enhance state-of-the-art methods with faster compression speed and more flexibility to partition an index. Experiments show that our methods offer a much better compression speed, while retaining an excellent space occupancy and decompression speed, networks.

A Flexible Space-Time Tradeoff on Hybrid Index with Bicriteria Optimization

Inverted indexes are widely adopted in the vast majority of information systems. Growing requirements for efficient query processing have motivated the development of various compression techniques with different spacetime characteristics. Although a single encoder yields a relatively stable point in the space-time tradeoff curve,flexibly transforming its characteristic along the curve to fit different information retrieval tasks can be a better way to prepare the index. Recent research comes out with an idea of integrating different encoders within the same index,namely, exploiting access skewness by compressing frequently accessed regions with faster encoders and rarely accessed regions with succinct encoders, thereby improving the efficiency while minimizing the compressed size.However, these methods are either inefficient or result in coarse granularity. To address these issues, we introduce the concept of bicriteria compression, which aims to formalize the problem of optimally trading the compressed size and query processing time for inverted index. We also adopt a Lagrangian relaxation algorithm to solve this problem by reducing it to a knapsack-type problem, which works in O(n log n)time and O(n)space, with a negligible additive approximation. Furthermore, this algorithm can be extended via dynamic programming pursuing improved query efficiency. We perform an extensive experiment to show that, given a bounded time/space budget, our method can optimally trade one for another with more efficient indexing and query performance.