INFOCOM 2026 Papers

As large language models are trained on datacenters with tens of thousands of compute nodes and are quickly becoming parts of our daily routines, the need for a flexible, easy-to-use, and high-performance research testbed for emulating and experimenting with new network protocols has become more pressing and relevant than ever. Conventional packet-level simulators are, by their nature of discrete-event simulation, not scalable enough; yet traditional network emulation testbeds, such as Mininet, are also showing their age with respect to the flexibility of implementing new algorithms, ability to run arbitrary application workloads, scalability to a large number of nodes, as well as the freedom of expanding beyond a single cluster to span multiple geographically distributed regions.In this paper, we present Nextmini, a modern, next-generation, high-performance networking research testbed for network emulation and experimentation. Implemented in Rust, it is designed from scratch to be as flexible as possible, accommodating a wider array of resource scheduling algorithms. It supports running arbitrary workloads — such as distributed machine learning training workloads — directly on the emulated network. Its design strikes an excellent balance between flexibility and performance, supporting both performant user-space emulation for maximum flexibility, as well as much higher kernel-level performance when users need such a performance boost. It is scalable to a larger number of nodes with ease in the same cluster, and can be easily expanded to span multiple geographically distributed datacenters. We conducted an extensive array of experiments to evaluate Nextmini’s tent-pole features, and to compare it with Mininet. Our results show both Nextmini’s raw power and its abundance of flexibility.

We consider a wireless scheduling problem in downlink wireless networks with unknown channel statistics. Scheduling performance relies heavily on accurate channel state information (CSI), which is often costly to acquire. In this study, CSI is obtained from ACK/NACK feedback, only after each scheduled transmission. Due to limited channel resources, all users cannot be scheduled for transmission simultaneously. Hence, the most recently observed CSI can be outdated. The traditional approach to solving scheduling problems using outdated CSI is to utilize belief states, which are calculated using the time correlation statistics of channels. However, channel statistics are often unknown; consequently, belief states can be uncountable and this approach becomes infeasible. In this paper, we introduce a new sufficient statistic-the latest observed CSI and its Age of Channel State Information (AoCSI), which characterizes the CSI staleness, to make the scheduling decisions. Accordingly, we are able to significantly reduce the state space. We develop an online Maximum Gain First (Online-MGF) policy which achieves sub-linear regret on the number of episodes. Numerical results demonstrate that Online-MGF policy converges to MGF and Whittle index policies with known channel statistics within a very few episodes. In addition, Online-MGF outperforms Maximum AoCSI First (MAF) and random policies.

Large language models (LLMs) enable personalized applications through parameter-efficient techniques like Low-Rank Adaptation (LoRA). Deploying LoRA-adapted LLMs faces a critical dilemma: resource-constrained user devices (UDs) are hard to host full-scale LLMs, while centralized edge deployments of all LoRA adapters incur huge storage and privacy risks. We propose EdgeLoRA, a collaborative framework where UDs retain private LoRA adapters locally while offloading base model computations to edge servers (ESs), enhancing privacy and reducing storage overhead. However, EdgeLoRA introduces latency inefficiency from multi-round ES-UD interactions and memory overhead from redundant per-user key-value (KV) caching. To address this, we design SemCache, a semantic-aware caching mechanism that exploits hierarchical redundancies. SemCache first clusters user queries into coarse-grained intents (e.g., health, economic) via lightweight textual encoders, then identifies reusable fine-grained subsequences (e.g., "feel pain in") within clusters. A hybrid-indexed global cache (cluster index + token position) enables cross-user KV reuse, while an adaptive policy optimizes cache utility via frequency, recency, and semantic impact metrics. Integrated with EdgeLoRA, SemCache reduces memory cost up to 10.8× and achieves 2.11× latency speedup per query via efficient token reuse, offering a scalable, privacy-preserving solution for efficient LoRA-adapted LLM inference at the edge.

Remote Direct Memory Access (RDMA) is increasingly adopted in container-based cloud environments for its ability to deliver high-performance communication, making it essential for distributed microservices and AI workloads. However, current container network interfaces (CNIs) face two critical limitations when supporting RDMA: (i) insufficient observability of RDMA operations and (ii) lack of transparent support for legacy TCP/IP containers. RDMA's kernel bypass nature prevents CNIs from monitoring container-level RDMA activities, hindering efficient resource monitoring and exposing multi-tenant environments to security risks. In addition, enabling RDMA for TCP/IP containers requires substantial code modifications due to the differences between RDMA verbs and socket APIs. This paper presents RDNet, a novel RDMA-aware CNI that integrates seamlessly with existing container platforms. RDNet leverages a user-space eBPF runtime to provide fine-grained RDMA observability at the container level and in-line RDMA policy enforcement with minimal overhead. For TCP/IP containers, RDNet relays traffic through a transparent RDMA proxy with eBPF-based traffic redirection, avoiding code changes while improving communication performance. Experimental results show RDNet incurs less than 5% monitoring overhead, blocks unauthorized RDMA operations, and enhances TCP/IP container throughput by up to 2× for HTTP and gRPC workloads.

Publicly available channel state information (CSI) in Wi-Fi has enabled a wide range of sensing applications and empowered integrated sensing and communication (ISAC). However, this same openness also raises serious privacy concerns. Existing defenses either neglect legitimate ISAC requirements or depend on costly, specialized hardware. To this end, we propose CloakFi, a plug-and-play, low-cost solution that leverages the channel manipulation capability of metasurfaces to safeguard privacy while preserving ISAC performance. The key idea is to assign each metasurface row two distinct beamforming vectors and randomly activate one at each time slot, introducing fluctuations that mask sensitive sensing information. Those vectors are jointly designed to remain nearly identical in the communication direction, ensuring stable, high-throughput transmission, while diverging in the sensing direction, thereby introducing significant fluctuations and obfuscating unauthorized observations. To further support legitimate sensing under such randomized configurations, we develop a time-domain masking and demasking method that associates CSI samples with their corresponding metasurface configurations and eliminates induced discrepancies, enabling accurate recovery of valid CSI. We implement CloakFi on commodity Wi-Fi devices, and extensive experiments demonstrate that it provides strong privacy protection without sacrificing communication or sensing performance for legitimate users.

Serverless inference is becoming increasingly popular in AI services thanks to its scalability and flexibility. However, high latency caused by pulling large container images and AI models remains a significant bottleneck. Lazy pulling, which pulls only the specific image layers and model layers when needed, helps reduce this latency by allowing the pulling process to overlap with inference execution in a pipeline manner. Nevertheless, we discover that lazy pulling exhibits interwoven dependencies between container image layers and AI model layers. This not only results in unnecessary idle periods where the system waits for the required layers to be pulled, but also complicates layer scheduling. To this end, in this paper, we investigate a Layer-wise Serverless Inference Scheduling (LSIS) problem, and formulate it as an Integer Linear Program (ILP) form. We further propose a LayerChain algorithm that utilizes a Variable-Length Time Slot (VL-Slot) and provides a theoretical analysis of its performance upper bound. Trace-driven evaluations demonstrate that our approach significantly reduces inference completion time compared with state-of-the-art methods by 38.49%.

Amplification Reflection Distributed Denial-of-Service (AR-DDoS) attacks remain a formidable threat, exploiting stateless protocols to flood victims with illegitimate traffic. Recent advances have enabled data-plane defenses against such attacks, but existing solutions typically assume symmetric routing and are limited to a single switch. These assumptions fail in modern networks where asymmetry is common, resulting in dropped legitimate responses and persistent connectivity issues. This paper presents ReAct, an in-network defense for AR-DDoS that is robust to asymmetry. ReAct performs request-response correlation across switches using programmable data planes and a sliding-window of Bloom filters. To handle asymmetric traffic, ReAct introduces a data-plane-based request forwarding mechanism, enabling switches to validate responses even when paths differ. ReAct can automatically adapt to routing changes with minimal intervention, ensuring continued protection even in dynamic network environments. We implemented ReAct on both a P4 interpreter and NVIDIA’s BlueField-3, demonstrating its applicability across multiple platforms. Evaluation results show that ReAct filters nearly all attack traffic without dropping legitimate responses-even under high-volume attacks and asymmetry. Compared to state-of-the-art approaches, ReAct achieves significantly lower false positives. To our knowledge, ReAct is the first data-plane AR-DDoS defense that supports dynamic, cross-switch collaboration, making it uniquely suitable for deployment in networks with asymmetry.

As Remote Direct Memory Access (RDMA) is applied in larger data centers, load balancing is more necessary to reduce congestion hotspots and handle failures. However, the existing load balancing solutions are hard to be deployed in practical data centers with heterogenous RDMA NICs (RNICs) since they either work in Linux kernel or require to modify RNICs. In this paper, we propose Tlaloc, a general-purpose load balancing solution with high performance and low resource consumption. We overcome the information and operational limitations caused by the lack of transmission logic on the host and make it compatible with heterogeneous RNICs. To this end, we make the following three efforts: 1) collect accurate network information in a timely manner through message splitting and pace control; 2) dynamically maintain a high-quality set of connections to resist uncertainty; 3) establish a consistent-pace load scheduling framework to achieve elastic load allocation. We deploy Tlaloc on commercial RNIC drivers. Testbed experiments show that Tlaloc can improve network utilization by up to 53% and improve the speed of distributed training by up to 34% compared to ECMP. In large-scale simulations, Tlaloc improves FCT by up to 20% compared to others under real workloads.

Predict-then-Optimize (PTO) pipelines are widely employed in computing and networked systems, where Machine Learning (ML) models are used to predict critical contextual information for downstream decision-making tasks such as cloud LLM serving, data center demand response, and edge workload scheduling. However, these ML predictors are often vulnerable to out-of-distribution (OOD) samples at test time, leading to significant decision performance degradation due to large prediction errors. To address the generalization challenges under OOD conditions, we present the framework of Distributionally Robust Decision-Focused Learning (DR-DFL), which trains ML models to optimize decision performance under the worst-case distribution. Instead of relying on classical Distributionally Robust Optimization (DRO) techniques, we propose Diffusion-Augmented Distributionally Robust Decision-Focused Learning (3D-Learning), which searches for the worst-case distribution within the parameterized space of a diffusion model. By lever-aging the powerful distribution modeling capabilities of diffusion models, 3D-Learning identifies worst-case distributions that remain consistent with real data, achieving a favorable balance between average and worst-case scenarios. Empirical results on an LLM resource provisioning task demonstrate that 3D-Learning outperforms existing DRO and Data Augmentation methods in OOD generalization performance.

We present VWAttacker, the first systematic testing framework for analyzing the security of Voice over WiFi (VoWiFi) User Equipment (UE) implementations. VWAttacker includes a complete VoWiFi network testbed that communicates with Commercial-Off-The-Shelf (COTS) UEs based on a simple interface to test the behavior of diverse VoWiFi UE implementations; uses property-guided adversarial testing to uncover security issues in different UEs systematically. To reduce manual effort in extracting and testing properties, we introduce an LLM-based, semi-automatic, and scalable approach for property extraction and testcase (TC) generation. These TCs are systematically mutated by two domain-specific transformations. Furthermore, we introduce two deterministic oracles to detect property violations automatically. Coupled with these techniques, VWAttacker extracts 63 properties from 11 specifications, evaluates 1,116 testcases, and detects 13 issues in 21 UEs. The issues range from enforcing a DH shared secret to 0 to supporting weak algorithms. These issues result in attacks that expose the victim UE's identity or establish weak channels, thus severely hampering the security of cellular networks. We responsibly disclose the findings to all the related vendors. At the time of writing, one of the vulnerabilities has been acknowledged by MediaTek with high severity.

Tor is a widely used network for anonymous communication, designed to safeguard the privacy and anonymity of both users and service providers through multi-layered encryption and multi-hop routing. Despite its robust architecture, Tor remains susceptible to various attacks, including denial-of-service (DoS) and deanonymization. However, these attacks are constrained by high resource requirements, scalability limitations, and the defenses implemented within the Tor network. Furthermore, they are particularly ineffective in identifying and circumventing the Guard nodes that protect onion services. In this paper, we uncover a novel vulnerability in Tor’s circuit construction process and bandwidth scheduling, termed the Circuit Circle vulnerability. Exploiting this flaw, attackers can create circular circuits, leading to bandwidth contention and overloading a Tor node. To demonstrate the severity of this vulnerability, we propose Duplicate-Node Attack, a three-phase strategy that identifies Guard nodes, exhausts their bandwidth, and performs coarse geolocation inference of onion services. Unlike conventional methods, our Guard node identification technique bypasses existing Tor Vanguard defenses without requiring control over any relays. Our extensive experimental results confirm that Duplicate-Node Attack can reliably identify Guard nodes, exhaust their bandwidth with minimal cost, and infer the geolocation of onion services, with an average latency bias of 35.4 ms.

Programmable networks are a revolutionary network paradigm that is widely deployed in modern data centers and enterprise networks. In programmable networks, topology discovery is responsible for building a global visibility of the data-plane topology, serving as the foundation of other key applications. However, existing topology discovery methods heavily rely on controllers to process massive LLDP packets, causing severe overhead that increases linearly with link changes. In this paper, we propose LTD, the first mechanism offloading core topology discovery to programmable data planes. LTD adopts three key strategies: three-layered maximum offloading that distributes tasks across switch ASIC, OS, and controller; an OS-ASIC co-driven mechanism that utilizes the characteristics of programmable switches to achieve periodic link change identification; and a plug-and-play data structure for seamless topology view updates. We conduct extensive evaluations on real-world topologies and hardware switches, and compare LTD with state-of-the-art methods. The results show that LTD achieves full link coverage, significantly reduces controller CPU consumption and TCP buffer occupancy (by 90% - 99%), and realizes high scalability (sublinear overhead increase with link changes).

System logs are central to diagnosing failures and monitoring behavior in complex software systems. Yet their unstructured format, semantic variability, and distributional shifts across systems make anomaly detection particularly challenging. Traditional methods—based on templates, static features, or rule-based matching—struggle to generalize, especially in low-label or cross-system scenarios. Large language models (LLMs) offer strong semantic understanding and can infer latent behavioral patterns from raw log text, while meta-learning enables fast adaptation to new tasks with few labels. To harness these complementary strengths, we propose a novel meta-learning framework that formulates log anomaly detection as a language-guided task induction problem. Rather than using LLMs as predictors or encoders, we employ them as semantic agents that decompose log data into structured meta-tasks and inform the adaptation of the meta-learner to new system environments. This synergy enables accurate and transferable anomaly detection across heterogeneous system environments, without requiring system-specific heuristics or extensive retraining. Experiments on diverse system log benchmarks demonstrate substantial gains in detection accuracy and robustness under distribution shift. To our knowledge, this is the first work to systematically integrate LLMs for task semantic modeling with meta-learning in system log anomaly detection, marking a shift toward generalizable, language-informed detection paradigms.

Tactile Internet (TI) requires ultra-low latency and high reliability to ensure stability and transparency in touch-enabled teleoperation. However, variable delays and packet loss present significant challenges to maintaining immersive haptic communication. To address this, we propose a predictive framework that integrates Discrete Mode Decomposition (DMD) with Shapley Mode Value (SMV) for accurate and timely haptic signal prediction. DMD decomposes haptic signals into interpretable intrinsic modes, while SMV evaluates each mode's contribution to prediction accuracy, which is well-aligned with the goal-oriented semantic communication. Integrating SMV with DMD further accelerates inference, enabling efficient communication and smooth teleoperation even under adverse network conditions.Extensive experiments show that DMD+SMV, combined with a Transformer architecture, outperforms baseline methods significantly. It achieves 98.9% accuracy for 1-sample prediction and 92.5% for 100-sample prediction, as well as extremely low inference latency: 0.056 ms and 2 ms, respectively. These results demonstrate that the proposed framework has strong potential to ease the stringent latency and reliability requirements of TI without compromising performance, highlighting its feasibility for real-world deployment in TI systems.

Palm vein recognition has gained traction in secure authentication due to its unique, stable, and inherently concealed biometric characteristics. However, the electromagnetic (EM) emissions from subcutaneous vein imaging sensors (SVIS) may unintentionally leak sensitive biometric information, fundamentally challenging the assumed security guarantees. In this paper, we propose VeinPhantom, a novel side-channel attack that reconstructs palm vein patterns from unintended EM emissions of SVIS, ultimately enabling spoofing attacks against biometric authentication systems. To overcome the challenge of low information entropy caused by weak EM signals and complex environmental interference, VeinPhantom first analyzes palm vein information from EM signals, then employs a cascaded enhancement strategy and incorporates a Dynamic Guidance Diffusion framework to progressively reconstruct high-fidelity palm vein patterns. Extensive experiments demonstrate that VeinPhantom achieves an average structure similarity index measure (SSIM) of 0.56 on commercial devices, along with a 56.96% spoofing success rate against state-of-the-art authentication systems. We further discuss potential mitigation strategies to defend against the attack.

Achieving intelligible and noise-robust voice communication during phone calls remains a significant challenge in real-world environments, where low signal-to-noise ratio (SNR) and background conversations are prevalent. While traditional audio-only speech enhancement systems show strong performance under moderate noise, they struggle in multi-speaker or low SNR scenarios due to the lack of user-specific cues. In this work, we propose LargeCall, a dual-modality speech enhancement framework that integrates microphone audio with inertial signals from the smartphone’s built-in accelerometer. Unlike visual or ultrasound-based methods, accelerometer sensing is passive, privacy-preserving, and robust in outdoor settings. Our design addresses three key challenges: (i) the scarcity of paired audio-accelerometer data is mitigated by bootstrapping from a pretrained audio-only encoder, reducing training data requirements and improving generalization; (ii) variability in phone orientation and body motion is handled through a posture-invariant transformation pipeline and a multi-dilated fusion module for robust feature extraction; and (iii) modality mismatch between audio and accelerometer streams is resolved via a cross-modal mask fusion strategy, which adaptively integrates complementary masks from both modalities without disrupting the pretrained encoder. We validate LargeCall on a cross-user evaluation protocol using realistic phone call scenarios. Experimental results show that LargeCall achieves substantial gains in both objective and subjective metrics across diverse noise conditions, demonstrating its effectiveness and real-world deployment potential. The demo of our system is available at https://anonymoususers718.github.io/largecall/.

Split learning recently emerged as a solution for distributed machine learning with heterogeneous IoT devices, where clients can offload part of their training to computationally-powerful helpers. The core challenge in split learning is to minimize the training time by jointly devising the client-helper assignment and the schedule of tasks at the helpers. We first study the model where each helper has a memory cardinality constraint on how many clients it may be assigned, which represents the case of homogeneous tasks. Through complexity theory, we rule out exact polynomial-time algorithms and approximation schemes even for highly restricted instances of this problem. We complement these negative results with a non-trivial polynomial-time 5-approximation algorithm. Building on this, we then focus on the more general heterogeneous task setting considered by Tirana et al. [INFOCOM 2024], where helpers have memory capacity constraints and clients have variable memory costs. In this case, we prove that, unless P = NP, the problem cannot admit a polynomial-time approximation algorithm for any approximation factor. However, by adapting our aforementioned 5-approximation algorithm, we develop a novel heuristic for the heterogeneous task setting and show that it outperforms heuristics from prior works through extensive experiments.