Umeå University's logo

The complete artifact for reproducing the results presented in the paper is provided. The artifact comprises source code, detection-model notebooks, an eBPF-collected system call dataset, and infrastructure scripts to recreate the testbed. It enables the reader to: (i) reproduce all detection experiments (SVM, XGBoost, 1D-CNN, Transformer, GNN) across four feature types (syscall identities, return values, parameters, and combined) and four context window sizes (250, 500, 1000, 2000) reported in Table V; (ii) reproduce the exploratory data analysis in Figures 4, 5(a), and 5(b); (iii) inspect the source code of all 30 serverless functions (Table III), including four attack payloads derived from real PyPI malware; (iv) reuse the system call dataset (~1.9 M traces) for independent research; and (v) reproduce the full testbed and data collection pipeline from scratch.

Mobile Edge Clouds (MECs) are emerging as a key complement to centralized cloud infrastructures by bringing computing and storage resources closer to the network edge, thereby reducing network bandwidth, latency, and jitter. A critical challenge in leveraging MECs effectively is the application placement problem, which seeks to minimize operational costs while ensuring end-user Quality of Service (QoS). This problem is further complicated by user mobility, as applications must migrate to maintain optimal QoS, yet frequent migrations can lead to unnecessary bandwidth consumption due to state transfer. In this paper, we tackle the application placement problem for stateful applications in MEC environments. We model the dynamic workloads, applications, and infrastructure typical of MECs and define the associated costs: resource utilization, migration, and QoS degradation. Based on this model, we propose two online placement algorithms – Gale-Shapley-based and Follow-me – designed to minimize the total cost of operating applications. These algorithms are compared against an offline benchmark that has complete future knowledge. Experimental results demonstrate that both proposed algorithms efficiently place applications in MECs, achieving operating costs within 8% of the global optimum approximated by the offline algorithm. Furthermore, the Gale-Shapley-based algorithm outperforms the Follow-me algorithm, reducing operating costs by up to 17% and improving load balancing across MECs to mitigate resource scarcity.

The rapid adoption of Artificial Intelligence (AI) in edge environments has accelerated the use of Function-as-a-Service (FaaS) serverless platforms for deploying workloads across the cloud-edge continuum. However, this architectural shift introduces critical security vulnerabilities either at individual clusters or in multi-cluster continuum systems. The AI functions frequently integrate unvetted open-source libraries, creating a supply-chain attack surface, where malicious dependencies can execute undetected inside short-lived containers and compromise those AI functions. Traditional static analysis often fails to uncover these dynamic threats, necessitating robust runtime analysis. To address these issues, we present a comprehensive evaluation of learning-based approaches for identifying compromised functions by analyzing the behaviors of system calls. Moreover, we generate a dataset of 1.6 million system calls from 30 different serverless functions deployed on an OpenFaaS-Kubernetes testbed, which is available to the community for benchmarking. Our systematic analysis and comparison of five learning-based detection methods across multiple feature representations provide further insights into the identification of compromised functions. Our results demonstrate that runtime system call analysis can effectively identify supply-chain compromises, achieving detection rates up to 100% with F1 scores of 99.90%, without requiring application or platform modifications. 

Web application services and networks encounter a broad range of security and performance anomalies, necessitating sophisticated detection strategies. However, performing anomaly detection in edge cloud environments, often constrained by limited resources, presents significant computational challenges and demands minimized detection time for real-time response. In this paper, we propose a model selection approach for resource efficient anomaly detection in edge clouds by leveraging an adapted Deep Q-Network (DQN) reinforcement learning technique. The primary objective is to minimize the computational resources required for accurate anomaly detection while achieving low latency and high detection accuracy. Through extensive experimental evaluation in our testbed setup over different representative scenarios, we demonstrate that our adapted DQN approach can reduce resource usage by up to 45 % and detection time by up to 85 % while incurring less than an 8 % drop in F1 score. These results highlight the potential of the adapted DQN model selection strategy to enable efficient, low-latency anomaly detection in resource-constrained edge cloud environments.

Edge intelligence enables the deployment of deep neural network (DNN) inference services across heterogeneous and resource-constrained edge servers. In such environments, inference requests must satisfy strict Quality-of-Service (QoS) requirements in terms of deadline and accuracy, while operating under limited computation, memory bandwidth, and energy budgets. The availability of multiple DNN model variants enables adaptive model selection but significantly increases the complexity of inference assignment decisions. Memory access rate plays a critical role in inference latency and energy consumption; however, existing model-selection approaches in this domain rarely incorporate memory bandwidth explicitly into both latency and energy models. In this work, we study QoS-aware DNN inference assignment in multi-server edge systems with energy constraints, and jointly model computation and memory bandwidth in both latency and energy. To address the NP-hard problem efficiently, we propose a polynomial-time, greedy, preference-based scheduling framework with two complementary assignment strategies, namely Model-oriented and Request-oriented. To analyze the performance of the proposed approaches, we conduct extensive experiments under different system settings using multiple DNN model families, along with comprehensive ablation studies. The results demonstrate the significant effect of DNN model characteristics, including computation, memory bandwidth, and output accuracy, as well as system configuration, on the performance of the strategies. Specifically, Model-oriented approaches achieve better performance for computation-intensive model variants or when inference service times are substantial. In contrast, Request-oriented approaches perform better for bandwidth-intensive model variants and shorter service times. We also demonstrate the notably lower run-time of the proposed Energy-aware-Model-oriented approach across all experiments, where such a lightweight and low-latency solution enables instantaneous decision making, critical for scalable, time-sensitive applications on resource-constrained edge servers.

Emerging deployments of Generative AI increasingly execute inference across decentralized and heterogeneous edge devices rather than on a single trusted server. In such environments, a single device failure or misbehavior can disrupt the entire inference process, making traditional best-effort peer-to-peer routing insufficient. Coordinating distributed generative inference therefore requires mechanisms that explicitly account for reliability, performance variability, and trust among participating peers.

In this paper, we present G-TRAC, a trust-aware coordination framework that integrates algorithmic path selection with system-level protocol design to ensure robust distributed inference.First, we formulate the routing problem as a Risk-Bounded Shortest Path computation and introduce a polynomial-time solution that combines trust-floor pruning with Dijkstra's search, achieving sub-millisecond median routing latency at practical edge scales, and remaining below 10 ms at larger scales.Second, to operationally support the routing logic in dynamic environments, the framework employs a Hybrid Trust Architecture that maintains global reputation state at stable anchors while disseminating lightweight updates to edge peers via background synchronization.

Experimental evaluation on a heterogeneous testbed of commodity devices demonstrates that G-TRAC significantly improves inference completion rates, effectively isolates unreliable peers, and sustains robust execution even under node failures and network partitions.

Machine learning models—deep neural networks in particular—have performed remarkably well on benchmark datasets across a wide variety of domains. However, the ease of finding adversarial counter-examples remains a persistent problem when training times are measured in hours or days and the time needed to find a successful adversarial counter-example is measured in seconds. Much work has gone into generating and defending against these adversarial counter-examples, however the relative costs of attacks and defences are rarely discussed. Additionally, machine learning research is almost entirely guided by test/train metrics, but these would require billions of samples to meet industry standards. The present work addresses the problem of understanding and predicting how particular model hyper-parameters influence the performance of a model in the presence of an adversary. The proposed approach uses survival models, worst-case examples, and a cost-aware analysis to precisely and accurately reject a particular model change during routine model training procedures rather than relying on real-world deployment, expensive formal verification methods, or accurate simulations of very complicated systems (e.g., digitally recreating every part of a car or a plane). Through an evaluation of many pre-processing techniques, adversarial counter-examples, and neural network configurations, the conclusion is that deeper models do offer marginal gains in survival times compared to more shallow counterparts. However, we show that those gains are driven more by the model inference time than inherent robustness properties. Using the proposed methodology, we show that ResNet is hopelessly insecure against even the simplest of white box attacks.

Accurate future workload prediction is an essential step for proactive resource allocation and efficient provisioning in cloud computing environments. Deep learning strategies have proven successful for this task, but they face challenges due to the high dimensionality of monitoring data, extensive preprocessing requirements, and computational overhead. In this paper, we propose a hybrid framework that integrates autoencoders for workload compression with Long Short-Term Memory (LSTM) networks for time-series forecasting. Unlike prior studies, our approach systematically analyzes the trade-off between compression ratio and predictive accuracy, demonstrating how dimensionality reduction can improve both scalability and robustness. Thereby reducing the computational burden associated with processing massive-scale monitoring data. Experiments conducted on both synthetic and real-world datasets demonstrate that the proposed method achieves up to 60% data compression with minimal reconstruction loss, while also improving prediction accuracy compared to baseline LSTM models. We evaluate the overall performance of the framework using various metrics, including data reduction ratio, prediction accuracy, and the effects of different compression stages on predictive performance. Additionally, we quantify the computational savings in terms of CPU usage, memory footprint, and training/inference times, confirming the framework's feasibility for real-world deployment. These results underscore the potential of integrating compression and prediction to achieve scalable, accurate, and resource-efficient management of cloud workloads.

Edge intelligence is about enabling deep learning applications to run on edge platforms, often under strict Quality of Service (QoS) constraints (e.g., deadlines and accuracy). The heterogeneity and limited computational and energy capacities of edge servers necessitate further study on energy-efficient Deep Neural Network (DNN) inference. While availability of DNN model variants enables adaptive selection without compromising accuracy, it increases the complexity of the solution space. Also, existing research on model selection for DNN inference lacks efficient estimation of energy consumption. This paper proposes a polynomial-time joint strategy for QoS-aware model instance provisioning and selection, based on many-to-many stable matching. Our novel formulation uses the number of floating-point operations (FLOPs) of each model along with hardware-level characteristics of edge servers to minimize total energy usage while maximizing successful inference completions. The proposed strategy is evaluated under different preference functions. Experimental results, compared to optimal and evolutionary algorithms, demonstrate the runtime efficiency of our strategy. Furthermore, extensive evaluations against baselines highlights its superior performance and the importance of jointly considering both system- and application-level parameters in the solution.

Recent advances in multi-core chip technology have enabled the dynamic tuning of shared memory resources, such as last-level cache and memory bus bandwidth. However, despite proven performance benefits, the complexity of effectively utilizing these hardware-level QoS enforcement features has limited their adoption in real-world cloud computing environments. In this paper, we introduce ESTHER, a novel approach to autonomously fine-tune QoS enforcement features in cloud environments using extremum seeking control, focusing on applications needs and operator ease-of-use. We demonstrate that ESTHER effectively maintains latency-critical workload SLOs and rapidly resolves any infringements by prioritizing shared memory resources. Such fast node-level resolution of SLO violations ensures that costly cluster-level scaling events may be avoided. Furthermore, ESTHER improves best-effort job throughput without impacting latency-critical workloads, achieving performance gains without utilizing workload profiling or prior knowledge of system dynamics.