The ACM CoNEXT 2025 Student Workshop (CoNEXT-SW'25) provides a focused venue for early-stage and ambitious research in computer networks and systems. The workshop complements the main conference by giving student authors the opportunity to present developing work in a constructive and technically rigorous setting. CoNEXT-SW'25 emphasizes clarity, reproducibility, and open discussion of methods and evaluation plans. This summary outlines the motivation, structure, and highlights of this year's workshop, which features novel contributions on digital twin synchronization, transport innovations in QUIC, hybrid load balancing, secure time synchronization, and machine learning-driven intrusion detection. Keynote: To be announced.
Artifact evaluation places a significant burden on both authors and reviewers: Authors must package bespoke experimental setups, while committees struggle to validate them. Research Infrastructures (RIs) offer tools to support reproducibility, yet interoperability remains limited due to insufficient standardization efforts. To address this, we propose new standards that unify RIs through a formal experiment execution format and topology management for networked experiments. In addition, we advance sustainability through integrated energy analysis and modeling. Together, these contributions improve reproducibility, enable cross-RI collaboration, and foster more sustainable experiments. This makes RIs more FAIR-particularly by strengthening interoperability and reusability-while aiding broader adoption across the (network) research community.
Ultra-wideband delivers superior indoor localization accuracy and is emerging as the nexus of modern industrial cyber-physical systems, catalyzing location-aware intelligence, rapid shop-floor reconfiguration, and real-time closed-loop motion control for mobile robotics. However, prevailing self-calibration approaches consistently fail when confronted with the demanding realities of industrial deployment, rendering theoretically sound methods operationally infeasible. We present an autonomous airborne self-calibration framework that seamlessly unifies network-geometry recovery and site-datum registration, realizing zero-skill staging, one-time drop-in deployment, zero-touch recalibration, and plant-scale operation.
The growth of commercial space enables applications such as realtime Earth observation and global edge computing. A natural question arises: can existing public cloud infrastructures be effectively extended to leverage LEO satellite links, providing scalable compute capabilities for satellite data processing and cross-region connectivity? Integrating public clouds with LEO networks, however, presents challenges such as dynamic satellite topologies, session continuity as visible satellites change, limited onboard resources, and underlay complexity for tenant VMs. To address this, we present IrisGW, a cloud gateway architecture enabling seamless cross-region VM communication over topology-varying satellites. Built atop mature cloud virtualization, IrisGW embeds destination locations into VXLAN tunnels for geo-based routing, allowing access satellites to forward packets efficiently. This reduces routing inflation, minimizes on-satellite state, and provides transparent access for tenant VMs. IrisGW demonstrates how public clouds and LEO networks can be coalesced to support robust, planet-scale cloud services.
The rise of programmable network devices - such as switches and SmartNICs - has enabled a new computing paradigm: In-Network Computing (INC). By performing computation directly in the data plane, INC can reduce communication overhead, accelerate distributed coordination, and improve the efficiency of system-wide operations. Distributed systems stand to benefit significantly from this paradigm shift. Many of their core functions-such as replication, consensus, scheduling, aggregation, and load balancing-are communication-heavy and latency-sensitive. In-network computing provides a compelling opportunity to rethink these mechanisms by pushing parts of their logic into the network fabric. While recent research has demonstrated the feasibility of INC for specific tasks (e.g., cache indexing, gradient aggregation, distributed counters), broader system-level integration remains challenging. Open questions span from programming abstractions and hardware constraints to performance modeling, security, and compatibility with evolving distributed runtimes. This workshop focuses on innetwork computing and AI for distributed systems (INCAS). We aim to bring together researchers and practitioners to share emerging ideas, explore new design patterns, and identify key challenges at the intersection of programmable networks and distributed system architecture.
IP addresses observed in Internet traffic exhibit rich cluster-within-cluster structure visually evocative of mathematical objects known as multifractals. Prior efforts to apply multifractal analysis to sets of IP addresses fell short of establishing robust statistical tools to demonstrate the existence and provide quantitative "fingerprints" of multifractal structure. Our work revitalizes multifractal analysis of IP addresses by developing a new analysis method rooted in robust statistics and tailored for several particular challenges that arise when applying such analysis to sets of IP addresses. To illustrate one potential usecase, we consider the problem of detecting "spatial" anomalies-i.e., points in time when the structure of traffic shifts in the address space-in unsampled netflow data collected on our campus network.
Low Earth Orbit (LEO) satellite networks suffer from frequent path changes due to satellite mobility, disrupting route stability and increasing control overhead. To address this, we propose a delay-bounded stable routing scheme with an adaptive threshold that reduces route updates while preserving low end-to-end (E2E) delay. Simulation results show about 45% fewer route updates with E2E delay kept within 5% of a traditional shortest-path routing scheme.
Edge computing orchestration faces significant challenges due to resource constraints, highly distributed topologies, and dynamic network conditions. The discrepancy between theoretical and actual runtime performance often leads to suboptimal deployment decisions. This discrepancy is severe in clustered deployments, as existing tools either saturate network links during testing or lack proactive assessment capabilities. None of these approaches accurately predicts service co-locating compatibility in real-world scenarios. We present KUT (Konnectivity Under Test), a lightweight network assessment framework designed specifically for edge environments, combining service-specific traffic simulation with periodical monitoring to provide accurate compatibility assessments without starving co-located services. KUT enables periodic background assessments that inform orchestration decisions while consuming minimal resources.
Large Language Model (LLM) training generates synchronized Remote Direct Memory Access (RDMA) bursts that heavily stress datacenter fabrics and are highly sensitive to faults. However, access to full-scale training clusters is costly, and existing network testers fail to accurately reproduce such patterns. We introduce GPTraffic, a topology- and model-aware testing framework that predicts, emulates, and analyzes LLM training workloads on programmable hardware. By combining burst-accurate traffic generation, RDMA-aware semantics, and fine-grained fault injection, GPTraffic enables scalable, realistic, and reproducible experiments that faithfully reflect the dynamics of distributed LLM training. This allows researchers to explore performance bottlenecks, congestion behavior, and fault tolerance under conditions that closely mirror real-world AI training workloads.
Queue management policies strongly influence latency, loss, and fairness, yet production switches often rely on static configurations. Changing policies in the wild is risky, as experiments can disrupt SLAs and queue dynamics are hard to reproduce offline. We present QueuePilot, a digital twin for switch queues that enables safe, traffic-driven evaluation of scheduling and Active Queue Management (AQM) policies. Our prototype, built on two Tofino switches, clones live traffic into the twin, multiplexes it across multiple queueing strategies, and monitors real-time fine-grained metrics. Preliminary results showthat QueuePilot can synchronize with the production queues and simultaneously evaluate alternative queuing policies, allowing precise real-time queue management without affecting live traffic.
Software vulnerabilities in widely deployed, Internet-exposed applications pose a significant threat, but they are only exploited at scale when attackers have a practical way to discover vulnerable instances. One factor that makes such a discovery easier may be Certificate Transparency (CT) logs: although designed to improve trust in digital certificates, they also expose information that attackers could leverage. In this work, we investigate the misuse of CT logs for large-scale reconnaissance of self-hosted web applications. By filtering a single day of CT logs with keywords from 27 popular web applications, we identify over 96 000 candidate domains. Crawling these domains reveals a median match rate of 10.9 %, with some instances exceeding 20 %. Our findings highlight that CT logs can provide attackers with an effective and timely method to identify potentially vulnerable instances of web applications, raising new considerations for Internet-scale security and privacy.
With the rapid advancement of Artificial Intelligence (AI), an increasing number of AI methods are being deployed on the edge for real-time analytics. However, due to the intensive computational demands of many AI workloads, especially limited by resource-constrained edge devices, they often need to be offloaded to the cloud. At the same time, applications with stringent privacy requirements necessitate that data be processed at the edge. This duality introduces new challenges for networking and system design within edge-cloud collaboration. Moreover, AI itself can serve as a powerful enabler to address the complexities of edge-cloud codeployment, optimizing resource allocation, data flow, and decisionmaking processes. ECCAI 2025 aims to explore this emerging space: a computing paradigm driven by AI and for AI in intelligent edgecloud collaboration. Specifically, the workshop investigates: Cloud-edge collaboration for AI: How to better meet the demands of real-time performance and strict privacy requirements in scenarios such as autonomous driving and smart healthcare; and how to more effectively leverage the edge for low-latency inference while utilizing the cloud for complex training and collaborative optimization. AI for cloud-edge collaboration: Leveraging AI to optimize the overall edge-cloud collaboration strategy, such as predicting workloads to decide whether tasks should be executed at the edge or in the cloud, and dynamically adjusting network bandwidth usage and compute resource allocation.
BlockNetSys promotes a network-aware co-design of consensus, dissemination, and verifiable networking. With 22 submissions, 10 were accepted ( 45.5%), spanning multicast propagation, eBPF/SGX-and SDN-based auditing, Bayesian/game-theoretic DoS defense, and ledger-driven decisions for edge/vehicular and extreme IoT. Results indicate that topology-aware propagation shortens confirmation tails and reduces traffic, tiered evidence (rich off-chain telemetry with on-chain commitments) enables line-rate, complianceready accountability, security posture must be tuned against liveness/finality, and edge learning gains from blockchain-backed provenance and incentives. We group contributions into four areas: Network-aware Consensus and Dissemination; On-/Off-chain Auditing and SDN Accountability; Security Modeling and DoS Defense; and Ledger-driven Edge/IoT Offloading and Coordination.
This talk examines the computing transformation during the era of modern AI revolution. Driven by high computing and bandwidth demand, scale-out datacenter architecture has shifted to scale-up super AI computers. And the evolution of AI models, from dense transformers, to sparse mixture-of-experts, to future agentic AI systems, continuously bring new types of workloads and cast new requirements to computing. Taken together, these render the next-generation accelerated and parallel techniques as well as distributed software designs.
Communication networks have evolved into critical infrastructure that underpins a vast range of consumer and industrial applications-from AR/VR and autonomous driving to smart manufacturing and telemedicine. Yet, current networking paradigms exhibit fundamental limitations in flexibility, scalability, service differentiation, mobility, security, and global operability, constraining their ability to meet emerging demands for intelligent, adaptive, and service-aware connectivity. The Polymorphic Network (PN) represents a next-generation architectural paradigm that unifies control and dynamic orchestration of forwarding behaviors and heterogeneous resources across the entire network. Building upon programmable networking foundations, PN enables the seamless coexistence and evolution of diverse networking paradigms (e.g., IP, NDN, MobilityFirst) while extending programmability into the computation, storage, and forwarding domains. This workshop invites contributions addressing the theory, design, implementation, applications, and operational insights of polymorphic networks, fostering interdisciplinary dialogue on how PN can drive the evolution of intelligent, secure, and adaptive future networks.
Integrated Satellite-Terrestrial Network (ISTN) suffers from packet losses caused by handovers and link interference, which conventional congestion control misinterprets as congestion signals and unnecessarily reduces congestion window (cwnd). We propose an Explicit Congestion Notification (ECN)-enabled congestion control mechanism with a Bayesian loss differentiation model that uses round-trip time (RTT) deviation and ECN-marked acknowledgements to identify congestion-induced losses.
Underwater acoustic networks are gaining importance across diverse applications, and simulation is a practical way to validate protocol designs given the high cost of real-world deployment. However, existing simulators suffer from low fidelity due to over-simplified abstractions and poor scalability in large networks. In this paper, we present a new simulation approach that executes the actual code designed for embedded Linux devices, thereby ensuring high fidelity. Our method intercepts system calls of real programs and integrates them with a parallel discrete-event simulator for improved efficiency. Preliminary results show that our approach achieves up to 4.8 Chi speedup over abstraction-based simulators for the example ALOHA protocol with 841 nodes.
Autonomous System (AS) relationships are the commercial agreements that define how ASes interact to route Internet traffic. For over a decade, researchers have relied on CAIDA's inferred AS relationships, which, however, lack validation against publicly available ground-truth data. This work presents the first systematic validation of CAIDA's AS relationships dataset using Autonomous System Provider Authorization (ASPA) objects - cryptographically signed records of the authorized providers of an AS. Analyzing 124 ASPA-compliant ASes, we find that CAIDA's inference aligns with ASPA data for only 22 ASes, often underestimating the number of providers and, in some cases, misclassifying them as customers.
Encrypted traffic has become the default in modern communication networks, yet its metadata often remains vulnerable to feature-based analysis and classification attacks. To address this, Differential Privacy (DP) based traffic obfuscation has been introduced as a defensive measure. Although it has demonstrated strong privacy guarantees, we observe that it inadvertently introduces distinctive statistical patterns that can themselves be exploited by adversaries, leading to detection or active blocking of the protected traffic. In this paper, we propose a hybrid defense framework that combines macro-level user behavior emulation with micro-level differential privacy mechanisms. At a high level, our system simulates realistic browsing and communication behaviors of ordinary users, while at a granular level, we inject DP-controlled random noise into traffic features. This joint design preserves user privacy without making the resulting flows statistically distinguishable from normal traffic.
The Internet's DNS and physical layers are critically co-dependent, yet this reliance is poorly quantified. We present a global measurement study of DNS reliance on the Submarine Cable Network (SCN). Using a novel geolocation-based inference method validated by physical constraints (speed-of-light), we analyze 12,827 traceroutes from 179 RIPE Atlas probes to root and TLD servers. Our key finding is that DNS reliance on SCN is significant but highly uneven. Root servers exhibit low reliance due to extensive anycast deployment, while TLD reliance varies widely, with Ghana's ccTLD reaching an alarming 92%. This demonstrates that anycast alone does not guarantee physical resilience and highlights critical vulnerabilities in the global DNS infrastructure.
