The rapid proliferation of Internet of Things (IoT) deployments across critical domains demands dependable systems that maintain reliability despite dynamic environmental conditions and resource constraints. Traditional approaches to ensuring data integrity, communication reliability, and security often assume independent and identically distributed (i.i.d.) data (concept drift), making them illsuited for the non-stationary nature of IoT environments. In practice, sensor faults, network partitions, and evolving attack behaviors frequently disrupt system performance, emphasizing the need for adaptive and resilient frameworks that can sustain trustworthy operation amid heterogeneous application demands, device capabilities, and network traffic dynamics. This research presents a cohesive framework of adaptive learning systems that collectively address these challenges across the IoT stack. Our contributions include energy-aware collaborative sensing with trust management, robust anomaly detection for noisy edge environments, fault-tolerant communication middleware, generative adaptation for network intrusion detection under concept drift, and continual learning for digital twins in non-stationary settings. Through extensive evaluation across multiple IoT domains, we demonstrate that these frameworks significantly enhance the dependability of data-driven IoT systems by enabling autonomous adaptation to changing conditions while operating under limited resource budgets.
Distributed machine learning faces fundamental challenges arising from heterogeneous data, shifting client populations, and volatile inference workloads. This dissertation advances the thesis that middleware-not the models themselves-must serve as an active, adaptive decision layer that unifies training-time and inference-time robustness. On the training side, FLIPS performs privacy-preserving participant selection via label-distribution clustering within TEEs, improving fairness and convergence under non-IID data, while ShiftEx detects covariate and label shifts in streaming environments and dynamically creates or reuses experts to accelerate adaptation with minimal retraining. On the inference side, OptiSeq improves the reliability of LLM-based API sequencing by selecting example orderings using inference-time log-probability reasoning, and VOILA formulates multimodal fidelity selection as a value-of-information optimization, achieving large cost reductions with calibrated accuracy guarantees. Together, these systems establish adaptive middleware as a unifying approach for improving efficiency, robustness, and accuracy across modern distributed ML deployments.
We demonstrate OptiSeq [1], an edge-cloud middleware system that converts natural language queries into executable API sequences with high reliability in distributed environments. A central challenge in such systems is that large language models (LLMs) can generate incorrect API sequences, causing costly retries that consume edge bandwidth and add cloud latency. Our demonstration shows how OptiSeq, a lightweight inference-time example-ordering optimization engine, reduces these failures by intelligently arranging in-context examples before prompting the LLM. In our distributed architecture, (1) natural language queries originate at the edge, (2) relevant examples are retrieved from an edge-hosted vector store, (3) OptiSeq runs within the middleware pipeline to optimize prompt construction, (4) a cloud-hosted LLM generates API sequences, and (5) a distributed API execution layer carries out the calls. Through live demonstrations on real-world API sequencing tasks, we show that OptiSeq-guided prompting reduces retry attempts by 35-45 % while maintaining high accuracy. This work presents the first demonstration of ordering-aware middleware for natural-language-to-API sequence generation deployed across edge-cloud systems.
Highly available applications rely on replicated data, but complex event interleavings between application logic and replicated data libraries (RDLs) often cause subtle integration bugs. Detecting such bugs is challenging due to the inherent nondeterminism of distributed execution, as certain bugs can only manifest under specific interleavings. Correctness testing, therefore, requires replaying all possible interleavings-a challenging task due to the combinatorial explosion of the interleaving space. My doctoral dissertation addresses this challenge with ER-pi, a middleware framework that exercises all possible interleavings between the application code and RDL; it also eliminates redundant and impossible interleavings via novel pruning techniques. Initial results show that ER-pi successfully reproduces 12 real-world bugs across multiple open-source RDLs while significantly reducing the interleaving search space. Our ongoing work extends this foundation with interleaving prioritization, ranking interleavings execution by their likelihood of exposing faults-particularly those introduced by recent code changes, thus accelerating bug discovery. This research supports developers responsible for ensuring the correctness and reliability of replicated data systems.
As institutions increasingly depend on networked systems, cyber-attacks pose growing risks, underscoring the need for autonomous defense agents to support human operators. Training such agents requires Autonomous Cyber Operation (ACO) Gyms-simulated environments where strategies can be learned and evaluated. Existing gyms like CybORG provide valuable baselines but lack realism. This demo paper presents DnD (Discover and Deploy), a model-driven engineering-based middleware for on-demand discovery and deployment of high-fidelity network topologies in both simulation and emulation. With DnD, users can scan, intuitively configure, and deploy realistic network scenarios in minutes. It features an OpenStack-based network scanner to infer topologies, a network importer to transform the scanned data into graphical models, and automated synthesis of Infrastructure-as-Code artifacts for deployment in RAMPART, the authors' agent training framework. Through a live use case, we demonstrate how DnD lowers the barrier to creating realistic and transferable ACO Gym environments-advancing the practicality of autonomous cyber defense.
Reducing Write Amplification (WAF) underpins SSD endurance. Flexible Data Placement (FDP) interfaces steer writes by lifetime/ hotness of data and enables targeted Garbage Collection (GC). However, broad, repeatable evaluation remains limited because physical FDP supported SSDs-particularly those exposing Reclaim Groups (RGs)-are scarce. FDPVirt+ is an NVMeVirt-based emulator that enables FDP interface with RG to control evaluations without the real devices. On mixed-hotness fio, enabling RG-based GC isolation yields up to 51% lowerWAF and 39% higher throughput versus a non-FDP baseline; at the filesystem level with F2FS, WAF drops by 9.7%. The platform offers an accessible, reproducible vehicle for FDP policy prototyping and app-level studies.
Federated Learning (FL) has emerged as an indispensable paradigm for privacy-preserving collaborative model training across distributed clients. Yet, conventional FL frameworks depend on a central server for aggregation, creating scalability limits, bias, and single-point vulnerabilities. Our previous work, FL-EGM, introduced a decentralized FL architecture that elects an aggregator in every round based on its validation accuracy. The elected aggregator abstains from local training, aggregates updates from other clients, and then refines the global model on its raw data, producing an Enhanced Global Model (EGM). Empirical evaluation yielded 98.54% accuracy and strong robustness against biased aggregators. This doctoral research extends FL-EGM toward a middleware-oriented decentralized FL framework that incorporates resource-aware client selection in adversarial training, energy-efficient scheduling, and incentive-driven participation. The goal is to build resilient, sustainable, and fair FL infrastructures capable of large-scale heterogeneous deployment.
Modern distributed systems replicate data across multiple execution sites by means of special-purpose replicated data libraries (RDLs), which provide read-write data access and synchronization. Programming languages often need to be mixed across replica sites to meet business requirements and resource constraints. Because RDLs are typically written in a single language, integrating them in multilingual environments requires special-purpose code, whose characteristics are poorly understood. We aim to bridge this knowledge gap by reviewing two key strategies for integrating RDLs in multilingual environments: (1) foreign-function interface (FFI) and (2) common data format (CDF). Our preliminary results indicate performance and implementation tradeoffs: CDF offers latency and memory consumption advantages, while incurring an additional implementation burden. With modern distributed systems utilizing multiple languages, our findings can inform the design of RDLs for multilingual replicated data systems.
Despite being widely accepted, serverless platforms provide significant overhead in terms of function execution and resource utilization. It has been demonstrated that serverless workloads often display bursts of concurrent function calls. These days, platforms don't handle such patterns, affecting performance significantly in terms of resource underutilization, execution time, and SLA violations. Also, individual functions exhibit diverse resource requirements. To overcome these challenges, we need intelligent techniques to provide better resource utilization and execution time. In this dissertation, first, I design Anubis, a fair and efficient resource scheduler for serverless applications, to meet the application demands in terms of resource, elasticity, performance, and cost. Second, I design Laius, an elastic and scalable scheduling framework for the ephemeral data management of serverless applications using Intel Optane Persistent Memory. Third, I intend to design a framework to accelerate the serverless directed acyclic graph (DAG) execution while optimizing the tail latency. Overall, my dissertation aims to focus on designing holistic resource management for serverless applications through performance modeling and systems development for higher performance and better resource utilization.
The emergence of Large Language Models (LLMs) has transformed software development, enabling automated coding for serverless computing. However, several challenges remain regarding the correctness, performance, cost-efficiency, and code quality of automatically generated functions. The proposed work first characterizes the variable nature of serverless code generation with respect to the key challenges. We then propose the creation of new agentic LLM agents to generate, optimize, and adapt serverless code to improve it's overall quality. We will innovate approaches to provide iterative feedback between LLMs and serverless platforms-for performance assessment, and static code analysis-for code quality assessment by leveraging LLM interoperability protocols. This research will ultimately contribute novel, performance-aware, and cross-platform generative AI tools for serverless computing.
As distributed systems grow in size, sophistication and complexity, designing and validating them at scale and under varied workload conditions to reason about their dependability and security properties before deploying them in production becomes even harder. Using limited, in-house resources for these purposes is infeasible thus requiring sophisticated playgrounds and testbeds. Although numerous testbeds have emerged to facilitate large-scale experimentation, each of these platforms focuses on specific technologies only and can thus help reason about only a small subset of the desired system properties. To get complete coverage forces the system designer to seek a range of testbeds that are often chosen arbitrarily, lack interoperability and require a steep learning curve all leading to very slow and suboptimal outcomes. This dissertation is investigating novel solutions utilizing the concept of Model-driven Middleware with support of Agentic/Generative AI techniques to overcome these challenges. In preliminary work, we have developed the Federated Experiments Design Studio (FEDS) framework that automates the federation of disparate testbeds and experiment orchestration and configuration. Multiple challenges still remain unresolved and solutions to these form the basis of this doctoral research.
Federated learning (FL) must support a client's right to be forgotten (a.k.a Federated Unlearning), enabling clients to have their data influence removed from a trained model. However, existing Federated Unlearning (FU) algorithms often incur high overhead, especially those based on influence functions that require computing secondorder Hessian matrices. We propose a novel algorithm to efficient FU that eliminates the need for any Hessian computation or inversion. Instead, our algorithm uses only first-order gradients to approximate each training point's influence via a first-order Taylor expansion. By avoiding expensive Hessian computations, we reduce the computational complexity of unlearning from cubic to roughly quadratic time (O (n(3)) to similar to O (n(2)) in the number of model parameters). Our algorithm operates locally on the client's data, preserving privacy by not involving other clients or raw data exchange.
Federated Learning (FL) enables a distributed client-server architecture where multiple clients collaboratively train a global Machine Learning (ML) model without sharing sensitive local data. However, FL often results in lower accuracy than traditional ML algorithms due to statistical heterogeneity across clients. Prior works attempt to address this by using model updates, such as loss and bias, from client models to select participants that can improve the global model's accuracy. However, these updates neither accurately represent a client's heterogeneity nor are their selection methods deterministic. We mitigate these limitations by introducing Terraform, a novel client selection methodology that uses gradient updates and a deterministic selection algorithm to hierarchically select heterogeneous clients for retraining. This bi-pronged approach allows Terraform to achieve up to 47% higher accuracy over prior works in our preliminary evaluation.
The leaderless consensus, also known as multi-leader consensus, drives the agreements on multiple blocks independently and simultaneously across instances, enhancing consensus throughput significantly. To resolve the conflicts between transactions, which access the same state, from various instances, a global ordering on these blocks is involved to establish a serializable execution sequence. However, this process suffers from the synchronization among instances, where the faster instances have to align with the delayed ones in each round, raising the overall commit latency. In this study, we propose MiAR-BFT, which achieves asynchronous running of instances by eliminating the round-based global ordering, thereby minimizing the impact of slower instances on overall performance. Additionally, MiAR-BFT exploits a synchronization mechanism to handle the execution of conflicting transactions, ensuring the global consistency without sacrificing the performance. The extensive evaluation confirms the advantage of MiAR-BFT over other state-of-the-art comparisons, reaching up to 3x higher throughput under network fluctuations or node imbalances.
With WebAssembly, you can write and run code in various languages on almost any platform. This has already led to its versatile use in IoT, edge, and cloud environments. With this broad use in complex distributed environments, additional control and management support, such as anomaly detection, to ensure reliable and secure execution will become key toWebAssembly's future success. We want to detect anomalies in WebAssembly modules to protect the system from bugs or malicious code. However, current anomaly detection solutions do not adhere to the WebAssembly philosophy by ignoring platform and source language independence or by being limited to specific types of attacks. In this paper, we presentWasmEye, a platform- and source languageindependent anomaly detection system for WebAssembly. WasmEye's anomaly detection is based on ensemble learning, directly integrated into the WebAssembly module. This allows WasmEye to offer platform-independent anomaly detection without any source language restrictions - a key feature for the WebAssembly ecosystem. We present the design and implementation of WasmEye, show how we can achieve secure anomaly detection inside of modules themselves and evaluate its performance characteristics and effectiveness of anomaly detection.
Conventional Function-as-a-Service (FaaS) systems provide limited support for the applications to configure FaaS deployments for performance needs, limiting the FaaS applicability and productivity. Recent work addresses these limitations by introducing a performance abstraction, which allows applications to specify their performance needs through predefined Software-level Agreements (SLAs). FaaS systems then use these SLAs to manage resources and scheduling, guaranteeing application performance. In this paper, we propose efficient methods for realizing a performance abstraction in FaaS systems. Our approach enables resource sharing across deployments and uses cloud resource allocation statistics to determine the optimal timing and quantity of resource provisioning. This allows us to meet SLAs cost-effectively while maintaining robustness across diverse applications and deployment environments. Our experiments demonstrate a 500x higher guarantee availability compared to current cloud standards, with 10x fewer resource costs. More extensive studies show that this efficiency is consistent across deployment scenarios, supporting up to 1,000 concurrent FaaS deployments with 99.95% availability and using less than 40% additional resources.
Trusted Execution Environments (TEEs) have been proposed as a promising approach for secure model inference, providing inuse data protection to ensure confidentiality and integrity against untrusted third parties, with additional attestability. However, TEEprotected secure model inference remains susceptible to numerous software vulnerabilities and fault attacks, potentially undermining the designed protection objectives and giving a false sense of security and reliability. To counter these diverse threats, we introduce MVTEE, a TEEbased model inference system employing Multi-Variant Execution (MVX) that runs multiple, diversified inference variants in parallel and monitors execution divergence at checkpoints. The idea of MVTEE is not to prevent threats, but to leverage the fact that a specific vulnerability typically impacts only one variant, ensuring timely threat detection and response before any damage is done. MVTEE applies a random-balanced model partitioning for checkpoint insertion and leverages the heterogeneous nature of model inference stack to generate variants with multi-level diversification. We base MVTEE on a cross-process monitoring architecture with a two-stage variant bootstrap design and support asynchronous selective MVX for efficient execution. Our evaluations demonstrate that MVTEE, in a real-world setup, secures model inference with acceptable performance overhead in sequential execution and maintains comparable or even improved performance in pipelined execution.
Modern data centers increasingly employ multi-tenant deployment models in which multiple applications or virtual machines share a single physical server. However, existing tiered memory management schemes classify pages solely by access frequency to govern promotions and demotions across memory tiers without accounting for the distinct access patterns of latency-critical (LC) and besteffort (BE) workloads. LC workloads demand low-latency service yet lack sustained high-frequency access; consequently, frequencybased tiering demotes LC data to slower memory (SMem), degrading responsiveness and violating service-level objectives (SLOs). To address these challenges, we propose MTAT, an adaptive tiered memory management framework that guarantees the SLO of LC workloads while maintaining overall system performance for BE workloads. Rather than relying solely on hotness-based page placement, MTAT employs distinct policies for LC and BE workloads by isolating them into dedicated fast memory (FMem) partitions. Specifically, MTAT employs reinforcement learning to identify the minimal FMem capacity necessary to satisfy stringent SLOs, supporting rapid response to sudden demand surges, and uses a simulated annealing algorithm to allocate the remaining FMem fairly among co-located BE workloads. Compared to state-of-theart tiered memory page-placement solutions, MTAT improves the maximum throughput of LC workloads by up to 1.7x and enhances BE workloads' fairness by up to 3.3x, all while incurring only a 19% throughput penalty at worst.
The Summary statistics are essential to analyse large datasets in various fields, including financial and medical research. Federated computations enhance statistical power by combining geo-distributed datasets while ensuring compliance with data protection regulations, privacy guarantees, and resilience against intrusions. We present Tides, a federated framework leveraging Trusted Execution Environments (TEEs) to defend against adversaries controlling up to.. of the.. datacenters. We present an instantiation of Tides using genomic (GWAS) statistics. We address TEE-specific attack vectors, including communication blocking and side-channel attacks. Tides follows the following three key steps: (1) TEEs share statistical results through reliable broadcast and run a randomized crash-tolerant binary consensus algorithm to identify the datasets that are available; (2) TEEs enforce differential privacy with ad hoc noise; and (3) TEEs run memory-oblivious algorithms to compute the final summary statistics. We implemented Tides with Intel SGX enclaves and demonstrated its practicality with three datasets.
