INFOCOM 2026 Papers

Training large language models (LLMs) typically involves a technique known as 3D parallelism, which combines Data Parallelism, Tensor Parallelism, and Pipeline Parallelism. While effective, this method necessitates frequent, simultaneous all-reduce communications across multiple dimensions among training nodes, leading to substantial network traffic and congestion. To mitigate these issues, this paper explores the integration of In-network Aggregation (INA) into 3D parallelism. This integration encounters challenges due to conflicts in multi-dimensional communication patterns, topology-induced band-width underutilization, and the need for dynamic adaptation. In response, we propose a novel network architecture, 3D-INA, which directly incorporates traffic patterns into hardware and replaces traditional all-reduce communications with INA. This approach effectively bridges the gap between INA and the traffic patterns of 3D parallelism, while also resolving mismatches in network topology. Additionally, we employ reconfigurable optical switches and adaptive placement strategies, enhancing the system’s adaptation to varying workloads. Our simulations on the ns-3 platform demonstrate that 3D-INA can increase training throughput by up to six times compared to traditional topologies and all-reduce algorithms, while reducing costs and increasing flexibility. Further evaluations on a P4 testbed confirm the practical feasibility of 3D-INA for real-world applications, offering a promising new architecture for efficient LLM training.

Localizing an odor-emitting source using distributed sensors is a fundamental challenge in autonomous sensing systems, with applications in environmental monitoring, pollution tracking, safety, and molecular communication. However, in practical implementations, centralized inference over high-resolution grids is computationally expensive and the complex dynamics of molecular dispersion pose significant challenges for reliable and efficient odor source localization (OSL). We propose a network-aware, resource-efficient solution for localizing the source in macro-scale environments where modeling of molecular concentration is governed by the advection-diffusion equation. Instead of computing the posterior/likelihood over the entire grid, we implement lightweight maximum a posteriori (MAP)/maximum likelihood estimation (MLE) methods using randomized candidate sampling. Each candidate location is evaluated through a PDE-based forward model guided by Gaussian prior centered on current estimate, and the best estimate guides the next round of sampling. This yields a convergent and cost-efficient localization strategy. We evaluate our method using synthetic experiments across various conditions. Results show convergence to the true source with decreasing estimation variance. We further evaluate estimator efficiency via the Cramér-Rao Lower Bound (CRLB) for this setting using finite-difference approximations and demonstrate that our method closely approaches this bound. Our open-source, modular simulation framework enables reproducible experiments and further research on OSL.

Large multimodal models (LMMs) are transforming traditional mobile video analytics into interactive services, where wearable cameras stream live videos and users pose free-form queries and receive responses in natural language. Due to the high deployment costs, LMMs are primarily deployed on cloud servers. Real-time video streaming under dynamic network conditions and the afterward inference thus significantly affect quality of experience (QoE). However, existing video analytics systems are designed for task-specific, frame-independent, single-modal tasks, making them unsuitable for interaction with LMMs over multimodal, contextual inputs. Even worse, the autoregressive decoding of LMMs delays visual token processing during text generation, affecting real-time response. To bridge these gaps, we present MoVi, the first collaborative system for real-time LMM-driven interactive video analytics over dynamic networks. MoVi first establishes an interaction-oriented QoE model tailored to emerging interactive video analytics applications based on real-world user studies. It then jointly designs the video streaming and inference stages for QoE optimization. Specifically, MoVi employs a causal-aware streaming controller to adapt video configurations under dynamic networks. A query-assisted token manager further reduces response latency by dynamically pruning buffered tokens. Extensive experiments on real-world datasets and user studies demonstrate that MoVi achieves 40.1% higher QoE and 38.9% higher user opinion scores than existing state-of-the-art systems.

In the field of network coding, the longstanding and fundamental multiple-unicast conjecture states that for independent unicast sessions in an undirected network, network coding offers no throughput advantage over routing. Despite its significance, this conjecture remains unresolved for over two decades, and has recently witnessed deep connections to core problems in computational complexity. Existing approaches, based on multicommodity flows, information inequalities, and geometric embeddings, have succeeded in verifying the conjecture for restricted network classes only.We introduce a new session dominance framework, whose core principle establishes that if the conjecture holds for a particular set of sessions, it must also hold for another, dominated set. This new perspective leads to: (1) unified proofs for known results; (2) a sufficient condition in the cost domain for the conjecture to hold on arbitrary undirected networks with three unicast sessions; (3) a new, equivalent formulation of the conjecture; and (4) a proof to the conjecture on a significant new network class defined solely by topological conditions—with no restrictions on network size or link capacities. The goal is to establish a new pathway toward resolving this important and longstanding open problem.

Network measurement is now a basic building block that supports network management applications to identify real-time events by measuring both large and small flows. However, existing techniques face the trade-off between high accuracy and resource efficiency: sketch-based techniques enjoy high accuracy and low resource consumption when measuring large flows but drop accuracy for small flows; in-band network telemetry (INT) measures every flow at the cost of high resource consumption. In this paper, we present MonPlan, a framework that utilizes both sketches and INT to measure large and small flows, respectively, offering both high measurement accuracy and resource efficiency. Unlike existing studies, our co-design is driven by two theoretical optimizations: (1) MonPlan adopts the near-optimal Lagrangian relaxation to solve the NP-hard selection of measurement points to deploy sketches and INT even when the routing information is unknown. (2) MonPlan estimates the worst-case rates of collecting measurement data. Its estimates support reinforcement learning to select paths without network congestions. We realize MonPlan on 12.8 Tbps programmable switches. With large-scale real-world topologies and nine network management applications, MonPlan provides 23∼95% higher accuracy and retains resource efficiency when compared with five types of existing techniques.

Environmental awareness is critical for maintaining robust connectivity in low Earth orbit (LEO) satellite networks. Starlink terminals construct a 123×123 obstruction map by passively recording signal quality from connected satellites. However, this process is slow: even under continuous operation, meaningful map coverage typically requires 6 to 12 hours, impractically slow for dynamic environments or short-term deployments. Building on Starlink’s existing capability, we investigate how intelligent satellite selection can further accelerate obstruction map construction, without requiring architectural changes. We formulate the problem as a multi-armed bandit with sleeping arms and introduce SkyBandit, a lightweight online learning algorithm that balances link quality with exploration of uncharted sky regions. SkyBandit achieves a regret bound of $O\left( {\sqrt {NT{\text{ln}}T} } \right)$, where N is the number of observed orbits over T rounds. Experiments using a refined version of xeoverse LEO emulator along with real Starlink measurement data demonstrate that SkyBandit reduces obstruction map convergence time to just around 3 hours, while preserving high-quality connections. To our knowledge, this is the first work to formalize and optimize obstruction map construction in LEO networks via principled sequential learning.

Network topology discovery is fundamental for understanding the structure of the Internet and for identifying security vulnerabilities. Although techniques such as Paris Traceroute have improved the reliability of path discovery, they remain challenging due to the limitations of the technique in adaptability and discovery effectiveness, particularly in terms of extensiveness and efficiency, especially as modern networks evolve toward highly dynamic and cloud-centric architectures. We present ATOPOS, the dynamic path exploration with adaptive probe construction method, to improve the above issues in large-scale topology discovery. First, ATOPOS selects the most responsive protocol and reliable fields to construct probes that adapt to the target network. Then, based on the adaptive probe, ATOPOS dynamically adjusts the probing path upper bound estimated by the discovered topology state and early stops by increment awareness pruning during the probing. At the same probing scale, ATOPOS discovers 1.36 times as many nodes and 2.98 times as many edges as state-of-the-art methods. We deploy ATOPOS across 29 Amazon regions to demonstrate its capability for uncovering cloud structures in large, dynamic network environments. We identify 4,060 critical nodes and 33,539 boundary routing nodes, and gather latency information to support deeper insights for future research.

Network protocols are foundational to modern communication systems, making vulnerability discovery critical. Stateful protocol fuzzers have been widely adopted due to their high efficiency and low false positives. However, they heavily depend on the quality of the initial seed corpus. Unfortunately, these seeds are typically handcrafted, limited in quantity, and often exercise only shallow protocol behaviors.In this paper, we present AutoSeeder, the first automated framework for constructing high-quality initial seed corpus tailored to stateful protocol fuzzing. Its core idea is to combine lightweight analysis of protocol source code with large language models to get state-handling logic and guide the generation of syntactically valid initial seeds that can reach deeper protocol states. Evaluation on twelve widely-used protocols shows that AutoSeeder can quickly construct a high-quality initial seed corpus, compared with four baselines, resulting in average improvements of 51.95% in state coverage, 152.6% in state transition coverage and 42.51% in code coverage. In addition, AutoSeeder is compatible with three state representation schemes to enhance the performance of the base fuzzers, including AFLNET, StateAFL, and NSFuzz. When integrated with AFLNET, AutoSeeder also discovers seven previously unknown bugs.

Bird’s-Eye-View (BEV) is critical to connected and automated vehicles (CAVs) as it can provide unified and precise representation of vehicular surroundings. However, quality of the raw sensing data may degrade in occluded or distant regions, undermining the fidelity of constructed BEV map. In this paper, we propose BEVCooper, a novel collaborative perception framework that can guarantee accurate and low-latency BEV map construction. We first define an effective metric to evaluate the utility of BEV features from neighboring CAVs. Then, based on this, we develop an online learning-based collaborative CAV selection strategy that captures the ever-changing BEV feature utility of neighboring vehicles, enabling the ego CAV to prioritize the most valuable sources under bandwidth-constrained vehicle-to-vehicle (V2V) links. Furthermore, we design an adaptive fusion mechanism that optimizes BEV feature compression based on the environment dynamics and real-time V2V channel quality, effectively balancing feature transmission latency and accuracy of the constructed BEV map. Theoretical analysis demonstrates that, BEVCooper achieves asymptotically optimal CAV selection and adaptive feature fusion under dynamic vehicular topology and V2V channel conditions. Extensive experiments on real-world testbed show that, compared with state-of-the-art benchmarks, the proposed BEVCooper enhances BEV perception accuracy by up to 63.18% and reduces end-to-end latency by 67.9%, with only 1.8% additional computational overhead.

Reconfigurable data center networks (DCNs) enhance traditional architectures with optical circuit switches (OCSs), enabling dynamic reconfiguration of inter-PoD links, i.e., the logical topology. Optimizing this topology is crucial for adapting to traffic dynamics but is challenging due to its combinatorial nature. The complexity increases further when demands can be distributed across multiple paths, requiring joint optimization of topology and routing. We propose Alternating Topology and Routing Optimization (ATRO), a unified framework that supports both one-hop topology optimization (where traffic is routed via direct paths) and multi-hop joint optimization (where routing is also optimized). Although these settings differ in constraints, both are combinatorially hard and challenge solver-based methods. ATRO addresses both cases efficiently: in the one-hop case, it guarantees the global optimum via an accelerated binary search; in the multi-hop case, it alternates between topology and routing updates, with routing steps optionally accelerated by existing traffic engineering (TE) methods. ATRO supports warm-starting and improves solution quality monotonically across iterations. ATRO remains competitive even when paired with solver-free TE methods, forming a fully solver-free optimization pipeline that still outperforms prior approaches in runtime and maximum link utilization across diverse workloads.

Federated fine-tuning of large language models (LLMs) enables collaborative tuning across distributed clients. However, due to the large size of LLMs, local updates in federated learning (FL) may incur substantial video random-access memory (VRAM) usage. Moreover, frequent model exchange may lead to significant communication overhead. To tackle these challenges, in this paper we propose ZorBA, a zeroth-order optimization-based federated fine-tuning framework with heterogeneous block activation. ZorBA leverages zeroth-order optimization to eliminate the storage of gradients at the clients by forward passes. ZorBA includes a heterogeneous block activation mechanism in which the central server allocates different subsets of transformer blocks to clients in order to accelerate the convergence rate and reduce the VRAM usage. Furthermore, ZorBA utilizes shared random seeds and the finite differences of gradients in order to reduce the communication overhead. We conduct theoretical analysis to characterize the effect of block activation decisions on the convergence rate and VRAM usage. To jointly enhance the convergence rate and reduce the VRAM usage, we formulate an optimization problem to optimize the block activation decisions. We propose an ϵ-constraint lexicographic algorithm to solve this problem. Experimental results show that ZorBA outperforms three federated fine-tuning baselines in VRAM usage by up to 62.41% and incurs a low communication overhead.

Home Internet-of-Things (H-IoT) automation is increasingly threatened by single-rule anomalies (SRA) and cross-rule interference (CRI). These threats can produce outcomes diverging from or even contradicting user expectations, compromising H-IoT system security and potentially endangering user safety and property. Existing detection techniques for these threats are fundamentally limited as they overlook undocumented command side effects, creating a critical blind spot in security analysis. This paper introduces Cet-Miner, a method that obtains and models these side effects through black-box command testing. Cet-Miner models system behavior as a Bayesian-smoothed Markov Decision Process (MDP) to characterize the non-deterministic side effects. For side effects that are not immediately observable, it uncovers them by comparing semantic inferences with test results, and represents them via a latent state structure introduced into the MDP. Building on this method, we propose HA-Inspector, an end-to-end threat detection system that implements both a runtime SRA monitor and a CRI model checker. We evaluate our method and system on a real-world H-IoT testbed comprising 38 devices from 9 vendors. Results show Cet-Miner discovers 5 distinct types of side effects, and HA-Inspector outperforms state-of-the-art tools in detecting both SRA and CRI threats.

Underwater sensing and communication are fundamental enablers for marine applications, such as autonomous exploration, environmental monitoring, and infrastructure inspection. However, conventional underwater communication techniques suffer from fundamental limitations: (i) acoustic-based methods offer long-range coverage but with limited data rates and high latency; (ii) radio-frequency signals suffer from severe attenuation in water, failing to provide communication for most underwater tasks; (iii) while Optical Wireless Communication (OWC) has emerged as a promising alternative due to its high bandwidth and low latency, its highly directional characteristic requires strict alignment between transceivers, posing significant challenges for Unmanned Underwater Vehicle (UUV) systems. To address these limitations, in this paper, we propose AquaSaC, a novel system that enables simultaneous underwater optical sensing and communication through a newly designed unified waveform modulation scheme. AquaSaC is inspired by the observation that frequency-modulated optical sensing signals in water can also be detected by optical receivers such as photodetectors. Building on this insight, we design AquaSaC with three key components: (i) Integrated Optical SensCom (IOSC) Modulator/Demodulator, (ii) Information-rich Beacon Generator (IBG), and (iii) Adaptive Waveform Optimization (AWO) module, with the objective to achieve precise sensing and high-speed communication simultaneously in dynamic underwater environments. To evaluate AquaSaC in practice, we design and implement a prototype using commercial off-the-shelf components and conduct extensive real-world underwater experiments to evaluate its sensing and communication performance. Results demonstrate that AquaSaC can achieve high-accuracy sensing, with an average error of ±5 cm in distance and ±10 cm/s in velocity estimation, while enabling reliable optical communication at Mbps-level data rates. These results highlight AquaSaC’s potential as a practical and efficient underwater integrated sensing and communication solution.

Magnetic Resonant Coupling Wireless Power Transfer (MRC-WPT) systems have emerged as a promising solution for contactless energy transmission. However, integrating data communication with power transfer remains challenging due to the conflicting requirements for efficiency and stability. In this work, we design and implement MiCC, a system that can simultaneously achieve communication and power supply using MRC-WPT signals. First, a carrier-modulated magnetic communication method is designed, which embeds information into the resonant signal to avoid interference with power control. Subsequently, a Binary Frequency Shift Keying (2FSK) modulation strategy based on frequency switching is introduced at the inverter level to enable data transmission. Finally, we develop a dual-carrier hierarchical modulation scheme that spectrally decouples power and communication signals, enhancing system robustness. Simulation results show that our system can achieve an information transmission rate of up to 20 kbps while maintaining a charging efficiency above 90%. Testbed experiments verified that, with minor modifications, data rates exceeding 18 kbps can be achieved on commercial off-the-shelf smartphones.

Collaborative perception enables Intelligent Connected Vehicles (ICVs) to overcome the limitations of individual sensors–such as occlusions and restricted sensing range–by enabling the exchange of perceptual data over vehicular networks. However, high-frequency V2V communication can saturate wireless bandwidth, cause channel contention, and degrade the freshness of safety-critical information. To address the fundamental challenge of determining the optimal timing, target, and content of V2V transmissions, we propose a Spatio-Temporal Dynamic Risk–based Collaborative Perception (STDR-CP) framework. This framework integrates Iterative Closest Point (ICP) registration with an Unscented Kalman Filter (UKF) to estimate the utility of sensed measurements under dynamic motion conditions. It further constructs spatiotemporal graphs with coupled interaction fields to forecast localized collision risks among surrounding vehicles. Guided by these insights, an adaptive scheduler selectively prioritizes and compresses high-value, high-risk data for broadcast, significantly reducing redundancy while preserving critical information. Extensive V2V simulations show that STDR-CP maintains end-to-end latency around 80ms and achieves over 92% demand coverage, while mitigating channel load and improving the timeliness of shared perception. These results show that risk-aware, content-adaptive scheduling can enable scalable, safety-critical perception in bandwidth-limited vehicular networks.

Emerging AI/ML techniques have been showing great potential in automating network control in open radio access networks (Open RAN). However, existing approaches heavily rely on blackbox policies parameterized by deep neural networks, which inherently lack interpretability, explainability, and transparency, and create substantial obstacles in practical network deployment. In this paper, we propose inRAN, a novel interpretable online Bayesian learning framework for network automation in Open RAN. The core idea is to integrate interpretable surrogate models and safe optimization solvers to continually optimize control actions, while adapting to non-stationary dynamics in real-world networks. We achieve the inRAN framework with three key components: 1) an interpretable surrogate model via ensembling Kolmogorov-Arnold Networks (KANs); 2) safe optimization solvers via integrating genetic search and trust-region descent method; 3) an online dynamics tracker via continual model learning and adaptive threshold offset. We implement inRAN in an end-to-end O-RAN-compliant network testbed, and conduct extensive over-the-air experiments with the focused use case of network slicing. The results show that, inRAN substantially outperforms state-of-the-art works, by guaranteeing the chance-based constraint with a 92.67% assurance ratio with comparative resource usage throughout the online network control, under unforeseeable time-evolving network dynamics.

Channel state information (CSI) acquisition and feedback overhead grows with the number of antennas, users, and reported subbands. This growth becomes a bottleneck for many antenna and reconfigurable intelligent surface (RIS) systems as arrays and user densities scale. Practical CSI feedback and beam management rely on codebooks, where beams are selected via indices rather than explicitly transmitting radiation patterns. Hardware-aware operation requires an explicit representation of the measured antenna/RIS response, yet high-fidelity measured patterns are high-dimensional and costly to handle. We present SPARK (Sparse Parametric Antenna Representation using Kernels), a training-free compression model that decomposes patterns into a smooth global base and sparse localized lobes. For 3D patterns, SPARK uses low-order spherical harmonics for global directivity and anisotropic Gaussian kernels for localized features. For RIS 1D azimuth cuts, it uses a Fourier-series base with 1D Gaussians. On patterns from the AERPAW testbed and a public RIS dataset, SPARK achieves up to 2.8× and 10.4× reductions in reconstruction MSE over baselines, respectively. Simulation shows that amortizing a compact pattern description and reporting sparse path descriptors can produce 12.65% mean uplink goodput gain under a fixed uplink budget. Overall, SPARK turns dense patterns into compact, parametric models for scalable, hardware-aware beam management.

Cloud networks are critical to modern services, but failures in these networks still happen frequently and are hard to diagnose. To improve reliability, we need to understand which parts of a network are most likely to fail and how different types of faults affect end-to-end communication. Fault injection (FI) is a useful way to study cloud network failures, but current FI tools focus on a single VM/container, APIs, or applications, and cannot test faults along the full network path from source to destination. To address this limitation, we propose CloudPathFI, a fault injection tool that takes actual network paths as input and orchestrates targeted fault injection campaigns along these paths. It supports 15 common fault types based on real cloud incident reports and can test faults across control, data, management, and physical layers. We evaluated CloudPathFI in 924 fault injection experiments, and the results show that CloudPathFI achieves full path coverage, far exceeding baseline tools (6%-60%), while maintaining high reliability (99.78% FI success rate) and low overhead (avg. 4.3% CPU), and uncovering several key findings, such as the fragility of virtual-physical boundaries.

Telecom Business Support Systems (TBSS) are critical to the operation and maintenance (O&M) of modern telecommunications networks. However, troubleshooting in TBSS is challenging due to diverse error descriptions, dependencies across multiple subsystems, and the constant emergence of new errors. Directly applying large language models (LLMs) to troubleshoot such ill-defined system problems remains difficult. This paper designs a knowledge-context augmented framework, KAT, that combines structured knowledge, context enhancement, and continual learning to unlock LLMs’ troubleshooting capabilities. Our KAT has been deployed in a commercial TBSS serving over 35 million users, handling more than 80k system errors monthly. It has also been successfully scaled to commercial TBSSs in 15 additional cities, automatically resolving 10k–20k system errors per month. After 10 months of testing, KAT achieves an average BERT-F1 of 86.53%, improves the monthly user satisfaction rate from 47% to 97%-99%, and reduces the average error duration from 51.4 to 5.9 hours, saving troubleshooting labor and improving user experience significantly.