INFOCOM 2023 Papers

Vertical Federated Learning (VFL) enables multiple data owners, each holding a different subset of features about largely overlapping sets of data sample(s), to jointly train a useful global model. Feature selection (FS) is important to VFL. It is still an open research problem as existing FS works designed for VFL either assumes prior knowledge on the number of noisy features or prior knowledge on the post-training threshold of useful features to be selected, making them unsuitable for practical applications. To bridge this gap, we propose the Federated Stochastic Dual-Gate based Feature Selection (FedSDG-FS) approach. It consists of a Gaussian stochastic dual-gate to efficiently approximate the probability of a feature being selected, with privacy protection through Partially Homomorphic Encryption without a trusted third-party. To reduce overhead, we propose a feature importance initialization method based on Gini impurity, which can accomplish its goals with only two parameter transmissions between the server and the clients. Extensive experiments on both synthetic and real-world datasets show that FedSDG-FS significantly outperforms existing approaches in terms of achieving accurate selection of high-quality features as well as building global models with improved performance.

Stochastic Multi-Armed Bandit (MAB) has recently been studied widely due to its vast range of applications. The classic model considers the reward of a pulled arm to be observed after a time delay that is sampled from a random distribution assigned for each arm. In this paper, we propose an extended framework in which pulling an arm gives both an instant (short-term) reward and a delayed (long-term) reward at the same time. The distributions of reward values for short-term and long-term rewards are related with a previously known relationship. The distribution of time delay for an arm is independent of the reward distributions of the arm. In our work, we devise three UCB-based algorithms, where two of them are near-optimal-regret algorithms for this new model, with the corresponding regret analysis for each one of them. Additionally, the random distributions for time delay values are allowed to yield infinite time, which corresponds to a case where the arm only gives a short-term reward. Finally, we evaluate our algorithms and compare this paradigm with previously known models on both a synthetic data set and a real data set that would reflect one of the potential applications of this model.

Recently, such industrial pioneers as Apple and Samsung have offered a new generation of offline finding network (OFN) that enables crowd search for missing devices without leaking private data. Specifically, OFN leverages nearby online finder devices to conduct neighbor discovery via Bluetooth Low Energy (BLE), so as to detect the presence of offline missing devices and report an encrypted location back to the owner via the Internet. The user experience in OFN is closely related to the success ratio (possibility) of finding the lost device, where the latency of the prerequisite stage, i.e., neighbor discovery, matters. However, the crowd-sourced finder devices show diversity in scan modes due to different power modes or different manufacturers, resulting in local optima of neighbor discovery performance. In this paper, we present a brand-new broadcast mode called ElastiCast to deal with the scan mode diversity issues. ElastiCast captures the key features of BLE neighbor discovery and globally optimizes the broadcast mode interacting with diverse scan modes. Experimental evaluation results and commercial product deployment experience demonstrate that ElastiCast is effective in achieving stable and bounded neighbor discovery latency within the power budget.

A fundamental challenge to providing edge-AI services is the need for a machine learning (ML) model that achieves personalization (i.e., to individual clients) and generalization (i.e., to unseen data) properties concurrently. Existing techniques in federated learning (FL) have encountered a steep tradeoff between these objectives and impose large computational requirements on edge devices during training and inference. In this paper, we propose SplitGP, a new split learning solution that can simultaneously capture generalization and personalization capabilities for efficient inference across resource-constrained clients (e.g., mobile/IoT devices). Our key idea is to split the full ML model into client-side and server-side components, and impose different roles to them: the client-side model is trained to have strong personalization capability optimized to each client’s main task, while the server-side model is trained to have strong generalization capability for handling all clients’ out-of-distribution tasks. We analytically characterize the convergence behavior of SplitGP, revealing that all client models approach stationary points asymptotically. Further, we analyze the inference time in SplitGP and provide bounds for determining model split ratios. Experimental results show that SplitGP outperforms existing baselines by wide margins in inference time and test accuracy for varying amounts of out-of-distribution samples.

5G positioning systems, as a solution for city-range integrated-sensing-and-communication (ISAC), are flooding into reality. However, the positioning security aspects of such an ISAC system have been overlooked, bringing threats to more than a billion 5G users. In this paper, we propose a new threat model for 5G positioning scenarios, namely the selective-PRS-spoofing attack (SPS), disabling the latest security enhancement method reported in 3GPP R18. In our attack pattern, the attacker first cracks the broadcast information of a 5G network and then poisons specific resource elements of the channel, which can introduce substantial localization errors at victims or even completely control the positioning results. Worse, such attacks are transparent to both the UE-end and the network-end due to their stealthiness and easily bypass the current 3GPP defense mechanisms. To solve this problem, a DL-based defense method called in-phase quadrature Network (IQ-Net) is proposed, which utilizes the hardware features of base stations to perform identification at the physical level, thereby thwarting SPS attacks on 5G positioning systems. Extensive experiments demonstrate that our method has 98% defense accuracy and good robustness to noise.

To perform advanced surveillance, Unmanned Aerial Vehicles (UAVs) require the execution of edge-assisted computer vision (CV) tasks. In multi-hop UAV networks, the successful transmission of these tasks to the edge is severely challenged due to severe bandwidth constraints. For this reason, we propose a novel A2-UAV framework to optimize the number of correctly executed tasks at the edge. In stark contrast with existing art, we take an application-aware approach and formulate a novel Application-Aware Task Planning Problem (A2-TPP) that takes into account (i) the relationship between deep neural network (DNN) accuracy and image compression for the classes of interest based on the available dataset, (ii) the target positions, (iii) the current energy/position of the UAVs to optimize routing, data pre-processing and target assignment for each UAV. We demonstrate A2-TPP is NP-Hard and propose a polynomial-time algorithm to solve it efficiently. We extensively evaluate A2-UAV through real-world experiments with a testbed composed by four DJI Mavic Air 2 UAVs. We consider state-of-the-art image classification tasks with four different DNN models (i.e., DenseNet, ResNet152, ResNet50 and MobileNet-V2) and object detection tasks using YoloV4 trained on the ImageNet dataset. Results show that A2-UAV attains on average around 38% more accomplished tasks than the state of the art, with 400% more accomplished tasks when the number of targets increase significantly. To allow full reproducibility, we pledge to share datasets and code with the research community.

Weak-signal detection underlies a variety of ubiquitous computing applications, such as wireless sensing and machinery fault diagnosis. Stochastic resonance (SR) provides a new way for weak-signal detection by boosting undetectable signals with added white noise. However, existing work has to take a long time to search optimal parameter settings for SR, which cannot fit well some time-critical applications. In this paper, we propose an adaptive SR scheme (ASR) that can amplify the original signal at a low cost in time. The basic idea is that we find that the potential parameter is a key factor that determines the performance of SR. By treating the system as a feedback loop, we can dynamically adjust the potential parameters according to the output signals and make SR happen adaptively. ASR answered two technical questions: how can we evaluate the output signal and how can we tune the potential parameters quickly towards the optimal. In ASR, we first design a spectral-analysis based solution to examine whether SR happens using continuous wavelet transform. After that, we reduce the parameter tuning problem to a constrained non-linear optimization problem and use the sequential quadratic programming to iteratively optimize the potential parameters. We implement ASR and apply it in two ubiquitous computing applications: respiration-rate detection and machinery fault diagnosis. Extensive experiments show that ASR outperforms the state-of-the-art.

Modern mmWave systems have limited scalability due to inflexibility in performing frequency multiplexing. All the frequency components in the signal are beamformed to one direction via pencil beams and cannot be streamed to other user directions. We present a new flexible mmWave system called mmFlexible that enables flexible directional frequency multiplexing, where different frequency components of the mmWave signal are beamformed in multiple arbitrary directions with the same pencil beam. Our system makes two key contributions: (1) We propose a novel mmWave front-end architecture called a delay-phased array that uses a variable delay and variable phase element to create the desired frequency-direction response. (2) We propose a novel algorithm called FSDA (Frequency-space to delay-antenna) to estimate delay and phase values for the real-time operation of the delay-phased array. Through evaluations with mmWave channel traces, we show that mmFlexible provides a 60-150% reduction in worst-case latency compared to baselines1.

Federated learning (FL) has recently emerged as a promising paradigm that trains machine learning (ML) models on clients' devices in a distributed manner without the need of transmitting clients' data to the FL server. In many applications of ML (e.g., image classification), the labels of training data need to be generated manually by human agents (e.g., recognizing and annotating objects in an image), which are usually costly and error-prone. In this paper, we study FL with crowdsourced data labeling where the local data of each participating client of FL are labeled manually by the client. We consider the strategic behavior of clients who may not make desired effort in their local data labeling and local model computation (quantified by the mini-batch size used in the stochastic gradient computation), and may misreport their local models to the FL server. We first characterize the performance bounds on the training loss as a function of clients' data labeling effort, local computation effort, and reported local models, which reveal the impacts of these factors on the training loss. With these insights, we devise Labeling and Computation Effort and local Model Elicitation (LCEME) mechanisms which incentivize strategic clients to make truthful efforts as desired by the server in local data labeling and local model computation, and also report true local models to the server. The truthful design of the LCEME mechanism exploits the non-trivial dependence of the training loss on clients' hidden efforts and private local models, and overcomes the intricate coupling in the joint elicitation of clients' efforts and local models. Under the LCEME mechanism, we characterize the server’s optimal local computation effort assignments and analyze their performance. We evaluate the proposed FL algorithms with crowdsourced data labeling and the LCEME mechanism for the MNIST-based hand-written digit classification. The results corroborate the improved learning accuracy and cost-effectiveness of the proposed approaches.

Federated Learning (FL) enables potentially a large number of clients to collaboratively train a global model with the coordination of a central cloud server without exposing client raw data. However, the FL model convergence performance, often measured by the job completion time, is hindered by two critical factors: non independent and identically distributed (non-IID) data across clients and the straggler effect. In this work, we propose a clustered FL framework, MCFL, to minimize the job completion time by mitigating the influence of non-IID data and the straggler effect while guaranteeing the FL model convergence performance. MCFL builds upon a two-stage operation: i) a clustering algorithm constructs clusters, each containing clients with similar computing and communications capabilities to combat the straggler effect within a cluster; ii) a deep reinforcement learning (DRL) algorithm based on soft actor-critic with discrete actions intelligently selects a subset of clients from each cluster to mitigate the impact of non-IID data, and derives the number of intra-cluster aggregation iterations for each cluster to reduce the straggler effect among clusters. Extensive testbed experiments are conducted under various configurations to verify the efficacy of MCFL. The results show that MCFL can reduce the job completion time by up to 70% compared with three state-of-the-art FL frameworks.

Traffic shaping is a common function at end hosts. Compared with hardware ones, software shapers are more flexible to be developed and deployed, and thus are very attractive. Nevertheless, software approaches are still unsatisfactory as they struggle to saturate 40Gbps and higher speed.While much effort has been made to reduce the intrinsic overhead of software traffic shaping, we find that it is the extrinsic overhead, such as PCIe communications and interrupts, that hinders shaping from achieving 40Gbps - 100Gbps speed. Batching is an effective way to amortize these overheads. However, blindly batching can degrade the network performance, as it introduces bursts into the network. Diving into the dilemma, we find that intra-flow burst is to blame for harming the network performance, while inter-flow burst, consisting of packets from different flows, can be naturally demultiplexed in the network.Based on the insight, we present FlowBundler, which can achieve efficient traffic shaping by inter-flow batching. Testbed experiments show that FlowBundler can achieve an accurate shaping of 98Gbps with a single CPU core, which is 2.6× better than state-of-the-art approaches. Large-scale simulations show that FlowBundler can batch packet transmissions without harming the network performance.

Federated Learning enables collaboratively model training among a number of distributed devices with the coordination of a centralized server, where each device alternatively performs local gradient computation and communication to the server. FL suffers from significant performance degradation due to the excessive communication delay between the server and devices, especially when the network bandwidth of these devices is limited, which is common in edge environments. Existing methods overlap the gradient computation and communication to hide the communication latency to accelerate the FL training. However, the overlapping can also lead to an inevitable gap between the local model in each device and the global model in the server that seriously restricts the convergence rate of learning process. To address this problem, we propose a new overlapping method for FL, AOCC-FL, which aligns the local model with the global model via calibrated compensation such that the communication delay can be hidden without deteriorating the convergence performance. Theoretically, we prove that AOCC-FL admits the same convergence rate as the non-overlapping method. On both simulated and testbed experiments, we show that AOCC-FL achieves a comparable convergence rate relative to the non-overlapping method while outperforming the state-of-the-art overlapping methods.

Fast reroute (FRR) mechanisms that can instantly handle network failures in the data plane are gaining attention in packet-switched networks. In FRR no notification messages are required as the nodes adjacent to the failure are prepared with a routing table such that the packets are re-routed only based on local information. However, designing the routing algorithm for FRR is challenging because the number of possible sets of failed network links and nodes can be extremely high, while the algorithm should keep track of which nodes are aware of the failure. In this paper, we propose a generic algorithmic framework that combines the benefits of Integer Linear Programming (ILP) and an effective approach from graph theory related to constructive graph characterization of k-connected graphs, i.e., edge splitting-off. We illustrate these benefits through arborescence design for FRR and show that (i) due to the ILP we have great flexibility in defining the routing problem, while (ii) the problem can still be solved very fast. We demonstrate through simulations that our framework outperforms state-of-the-art FRR mechanisms andvprovides better resilience with shorter paths in the arborescences.

The programmable switch offers a limited capacity of packet header vector (PHV) words that store packet header fields and metadata fields defined by network functions. However, existing switch compilers employ inefficient strategies of encoding fields on PHV words. Their encoding wastes scarce PHV words and may result in failures when deploying network functions. In this paper, we propose Melody, a new framework that reuses PHV words for as many fields as possible to achieve resource-efficient PHV encoding. Melody offers a field analyzer and an optimization framework. The analyzer identifies which fields can reuse PHV words while preserving the original packet processing logic. The framework integrates analysis results into its encoding to offer the resource-optimal decisions. We evaluate Melody with production-scale network functions. Our results show that Melody reduces the consumption of PHV words by up to 85%.