INFOCOM 2023 Papers

Multicast traffic is growing rapidly due to the development of multimedia streaming. Lately, stateless multicast protocols, such as BIER, have been proposed to solve the excessive routing states problem of traditional multicast protocols. However, the high complexity of multicast tree computation and the limited scalability for concurrent requests still pose daunting challenges, especially under dynamic group membership. In this paper, we propose Hawkeye, a dynamic and stateless multicast mechanism with deep reinforcement learning (DRL) approach. For real-time responses to multicast requests, we leverage DRL enhanced by a temporal convolutional network (TCN) to model the sequential feature of dynamic group membership and thus is able to build multicast trees proactively for upcoming requests. Moreover, an innovative source aggregation mechanism is designed to help the DRL agent converge when faced with a large amount of multicast requests, and relieve ingress routers from excessive routing states. Evaluation with real-world topologies and multicast requests demonstrates that Hawkeye adapts well to dynamic multicast: it reduces the variation of path latency by up to 89.5% with less than 12% additional bandwidth consumption compared with the theoretical optimum.

Traffic Engineering (TE) has been widely used by network operators to improve network performance and provide better service quality to users. One major challenge for TE is how to generate good routing strategies adaptive to highly dynamic future traffic scenarios. Unfortunately, existing works could either experience severe performance degradation under unexpected traffic fluctuations or sacrifice performance optimality for guaranteeing the worst-case performance when traffic is relatively stable. In this paper, we propose LARRI, a learning-based TE to predict adaptive routing strategies for future unknown traffic scenarios. By learning and predicting a routing to handle an appropriate range of future possible traffic matrices, LARRI can effectively realize a trade-off between performance optimality and worst-case performance guarantee. This is done by integrating the prediction of future demand range and the imitation of optimal range routing into one step. Moreover, LARRI employs a scalable graph neural network architecture to greatly facilitate training and inference. Extensive simulation results on six real-world network topologies and traffic traces show that LARRI achieves near-optimal load balancing performance in future traffic scenarios with up to 43.3% worst-case performance improvement over state-of-the-art baselines, and also provides the lowest end-to-end delay under dynamic traffic fluctuations.

Decentralized bilevel optimization has received increasing attention recently due to its foundational role in many emerging multi-agent learning paradigms (e.g., multi-agent meta-learning and multi-agent reinforcement learning) over peer-to-peer edge networks. However, to work with the limited computation and communication capabilities of edge networks, a major challenge in developing decentralized bilevel optimization techniques is to lower sample and communication complexities. This motivates us to develop a new decentralized bilevel optimization called DIAMOND (decentralized single-timescale stochastic approximation with momentum and gradient-tracking). The contributions of this paper are as follows: i) our DIAMOND algorithm adopts a single-loop structure rather than following the natural double-loop structure of bilevel optimization, which offers low computation and implementation complexity; ii) compared to existing approaches, the DIAMOND algorithm does not require any full gradient evaluations, which further reduces both sample and computational complexities; iii) through a careful integration of momentum information and gradient tracking techniques, we show that the DIAMOND algorithm enjoys $\mathcal{O}\left( {{ \in ^{ - 3/2}}} \right)$ in sample and communication complexities for achieving an ϵ-stationary solution, both of which are independent of the dataset sizes and significantly outperform existing works. Extensive experiments also verify our theoretical findings.

Thanks to regulatory policies such as the General Data Protection Regulation (GDPR), it is essential to provide users with the right to erasure regarding their own private data, even if such data has been used to train a neural network model. Such a machine unlearning problem becomes even more challenging in the context of federated learning, where clients collaborate to train a global model with their private data. When a client requests its data to be erased, its effects have already gradually permeated through a large number of clients, as the server aggregates client updates over multiple communication rounds. All of these affected clients need to participate in the retraining process, leading to prohibitive retraining costs with respect to the wall-clock training time.In this paper, we present the design and implementation of Knot, a new clustered aggregation mechanism custom-tailored to asynchronous federated learning. The design of Knot is based upon our intuition that, with asynchronous federated learning, clients can be divided into clusters, and aggregation can be performed within each cluster only so that retraining due to data erasure can be limited to within each cluster as well. To optimize client-cluster assignment, we formulated a lexicographical minimization problem that could be transformed into a linear programming problem and solved efficiently. Over a variety of datasets and tasks, we have shown clear evidence that Knot outperformed the state-of-the-art federated unlearning mechanisms by up to 85% in the context of asynchronous federated learning.

Transforming off-the-shelf deep neural network (DNN) models into dynamic multi-exit architectures can achieve inference and transmission efficiency by fragmenting and distributing a large DNN model in edge computing scenarios (e.g., edge devices and cloud servers). In this paper, we propose a novel backdoor attack specifically on the dynamic multi-exit DNN models. Particularly, we inject a backdoor by poisoning one DNN model's shallow hidden layers targeting not this vanilla DNN model but only its dynamically deployed multi-exit architectures. Our backdoored vanilla model behaves normally on performance and cannot be activated even with the correct trigger. However, the backdoor will be activated when the victims acquire this model and transform it into a dynamic multi-exit architecture at their deployment. We conduct extensive experiments to prove the effectiveness of our attack on three structures (ResNet-56, VGG-16, and MobileNet) with four datasets (CIFAR-10, SVHN, GTSRB, and Tiny-ImageNet) and our backdoor is stealthy to evade multiple state-of-the-art backdoor detection or removal methods.

IEEE INFOCOM 2023 - IEEE Conference on Computer Communications, New York City, NY, USA, 17-20 May 2023

Cardinality estimation is a fundamental problem with diverse practical applications. HyperLogLog (HLL) has become a standard in practice because it offers good memory efficiency, constant update time, and mergeability. Some recent work achieved better memory efficiency, but typically at the cost of impractical update time or losing mergeability, making them incompatible with applications like network-wide traffic measurement. This work presents SpikeSketch, a better cardinality estimator that reduces memory usage of HLL by 37% without sacrificing other crucial metrics. We adopt a bucket-based data structure to promise constant update time, design a smoothed log4ranking and a spike coding scheme to compress cardinality observables into buckets, and propose a lightweight mergeable lossy compression to balance memory usage, information loss, and mergeability. Then we derive an unbiased estimator for recovering cardinality from the lossy-compressed sketch. Theoretical and empirical results show that SpikeSketch can work as a drop-in replacement for HLL because it achieves a near-optimal MVP (memory-variance-product) of 4.08 (37% smaller than HLL) with constant update time and mergeability. Its memory efficiency even defeats ACPC and HLLL, the state-of-the-art lossless-compressed sketches using linear-time compression to reduce memory usage.

Introducing the emerging serverless paradigm into edge computing could avoid over- and under-provisioning of limited edge resources and make complex edge resource management transparent to application developers, which largely facilitates the cost-effectiveness, portability, and short time-to-market of edge applications. However, the computation/data dispersion and device/network heterogeneity of edge environments prevent current serverless computing platforms from acclimating to the network edge. In this paper, we address such challenges by formulating a container placement and data flow routing problem, which fully considers the heterogeneity of edge networks and the overhead of operating serverless platforms on resource-limited edge servers. We design an online algorithm to solve the problem. We further show its local optimum for each arriving container and prove its theoretical guarantee to the optimal offline solution. We also conduct extensive simulations based on practical experiment results to show the advantages of the proposed algorithm over existing baselines.

We consider the fundamental problem of designing a self-adjusting tree, which efficiently and locally adapts itself towards the demand it serves (namely accesses to the items stored by the tree nodes), striking a balance between the benefits of such adjustments (enabling faster access) and their costs (reconfigurations). This problem finds applications, among others, in the context of emerging demand-aware and reconfigurable datacenter networks and features connections to self-adjusting data structures. Our main contribution is SeedTree, a dynamically optimal self-adjusting tree which supports local (i.e., greedy) routing, which is particularly attractive under highly dynamic demands. SeedTree relies on an innovative approach which defines a set of unique paths based on randomized item addresses, and uses a small constant number of items per node. We complement our analytical results by showing the benefits of SeedTree empirically, evaluating it on various synthetic and real-world communication traces.

Towards addressing spectral scarcity and enhancing resource utilization in 5G networks, network slicing is a promising technology to establish end-to-end virtual networks without requiring additional infrastructure investments. By leveraging Software Defined Networks (SDN) and Network Function Virtualization (NFV), we can realize slices completely isolated and dedicated to satisfy the users' diverse Quality of Service (QoS) prerequisites and Service Level Agreements (SLAs). This paper focuses on the technical and economic challenges that emerge from the application of the network slicing architecture to real-world scenarios. We consider a market where multiple Network Providers (NPs) own the physical infrastructure and offer their resources to multiple Service Providers (SPs). Then, the SPs offer those resources as slices to their associated users. We propose a holistic iterative model for the network slicing market along with a clock auction that converges to a robust $ε$-competitive equilibrium. At the end of each cycle of the market, the slices are reconfigured and the SPs aim to learn the private parameters of their users. Numerical results are provided that validate and evaluate the convergence of the clock auction and the capability of the proposed market architecture to express the incentives of the different entities of the system.

Graph Neural Networks (GNNs) are increasingly popular due to excellent performance on learning graph-structured data in various domains. With fast expanding graph sizes and feature dimensions, distributed GNN training has been adopted, with multiple concurrent workers learning on different portions of a large graph. It has been observed that a main bottleneck in distributed GNN training lies in graph feature fetching across servers, which dominates the training time of each training iteration at each worker. This paper studies efficient feature caching on each worker to minimize feature fetching overhead, in order to expedite distributed GNN training. Current distributed GNN training systems largely adopt static caching of fixed neighbor nodes. We propose a novel two-level dynamic cache design exploiting both GPU memory and host memory at each worker, and design efficient two-level dynamic caching algorithms based on online optimization and a lookahead batching mechanism. Our dynamic caching algorithms consider node requesting probabilities and heterogeneous feature fetching costs from different servers, achieving an O(log3k) competitive ratio in terms of overall feature-fetching communication cost (where k is the cache capacity). We evaluate practical performance of our caching design with testbed experiments, and show that our design achieves up to 5.4x convergence speed-up.

When congestion control algorithms compete on shared links, unfair outcomes can result, especially between algorithms that aim to prioritize different objectives. For example, a throughput-maximizing application could make the link completely unusable for a latency-sensitive application. In order to study these outcomes formally, we model the congestion control problem as a game in which applications have heterogeneous utility functions. We draw on the comparative statics literature in economics to derive simple and practically useful conditions under which all applications achieve at least ε utility at equilibrium, a minimal safety condition for the network to be useful for any application. Compared to prior analyses of similar games, we show that our framework supports a more realistic class of utility functions that includes highly latency-sensitive applications such as teleconferencing and online gaming.

Despite the substantial success of deep learning for Automatic Modulation Classification (AMC), models trained on a specific transmitter configuration and channel model often fail to generalize well to other scenarios with different transmitter configurations, wireless fading channels, or receiver impairments such as clock offset. This paper proposes Contrastive Learning with Self-Reconstruction called CLSR-AMC to learn good representations of signals resilient to channel changes. While contrastive loss focuses on the differences between individual modulations, the reconstruction loss captures representative features of the signal. Additionally, we develop three data augmentation operators to emulate the impact of channel and hardware impairments without exhaustive modeling of different channel profiles. We perform extensive experimentation with commonly used realistic datasets. We show that CLSR-AMC outperforms its counterpart based on contrastive learning for the same amount of labeled data by significant average accuracy gains of 24.29%, 17.01%, and 15.97% in the Additive White Gaussian Noise (AWGN), Rayleigh, and Rician channels, respectively.

In this paper we investigate the limit performance of Floating Gossip, a new, fully distributed Gossip Learning scheme which relies on Floating Content to implement location-based probabilistic evolution of machine learning models in an infrastructure-less manner. We consider dynamic scenarios where continuous learning is necessary, and we adopt a mean field approach to investigate the limit performance of Floating Gossip in terms of amount of data that users can incorporate into their models, as a function of the main system parameters. Different from existing approaches in which either communication or computing aspects of Gossip Learning are analyzed and optimized, our approach accounts for the compound impact of both aspects. We validate our results through detailed simulations, proving good accuracy. Our model shows that Floating Gossip can be very effective in implementing continuous training and update of machine learning models in a cooperative manner, based on opportunistic exchanges among moving users.

This paper studies an online service caching problem, where an edge server, equipped with a prediction window of future service request arrivals, needs to decide which services to host locally subject to limited storage capacity. The edge server aims to minimize the sum of a request forwarding cost (i.e., the cost of forwarding requests to remote data centers to process) and a service instantiating cost (i.e., that of retrieving and setting up a service). Considering request patterns are usually non-stationary in practice, the performance of the edge server is measured by dynamic regret, which compares the total cost with that of the dynamic optimal offline solution. To solve the problem, we propose a randomized online algorithm with low complexity and theoretically derive an upper bound on its expected dynamic regret. Simulation results show that our algorithm significantly outperforms other state-of-the-art policies in terms of the runtime and expected total cost.

The emerging digital twin technique enhances the network management efficiency and provides comprehensive insights, through mapping physical objects to their digital twins. The user satisfaction on digital twin-enabled query services relies on the freshness of digital twin data, which is measured by the Age of Information (AoI). Because the remote cloud faces challenges in providing data for users due to long service delays, Mobile Edge Computing (MEC), as a promising technology, offers real-time data communication between physical objects and their digital twins at the edge of the core network. However, the mobility of physical objects and dynamic query arrivals make efficient service provisioning in MEC become challenging. In this paper, we investigate the dynamic digital twin placement for improving user service satisfaction in MEC environments. We focus on two user service satisfaction augmentation problems under both static and dynamic digital twin placement schemes: the static and dynamic utility maximization problems. We first formulate an Integer Linear Programming (ILP) solution to the static utility maximization problem when the problem size is small; otherwise, we propose a performance- guaranteed approximation algorithm for it. We then devise an online algorithm for the dynamic utility maximization problem with a provable competitive ratio. Finally, we evaluate the performance of the proposed algorithms through experimental simulations. Simulation results demonstrate that the proposed algorithms outperform the comparison baseline algorithms, and the performance improvement is no less than 11.6%, compared with the baseline algorithms.

We propose Argosleep, a millimeter-wave (mmWave) wireless sensors based sleep posture monitoring system that predicts the 3D location of body joints of a person during sleep. Argosleep leverages deep learning models and knowledge of human anatomical features to solve challenges with low-resolution, specularity, and aliasing in existing mmWave devices. Argosleep builds the model by learning the relationship between mmWave reflected signals and body postures from thousands of existing samples. Since practical sleep also involves sudden toss-turns, which could introduce errors in posture prediction, Argosleep designs a state machine based on the reflected signals to classify the sleeping states into rest or toss-turn, and predict the posture only during the rest states. We evaluate Argosleep with real data collected from COTS mmWave devices for 8 volunteers of diverse ages, gender, and height performing different sleep postures. We observe that Argosleep identifies the toss-turn events accurately and predicts 3D location of body joints with accuracy on par with the existing vision-based system, unlocking the potential of mmWave systems for privacy-noninvasive at-home healthcare applications.