Συλλογή και ανάλυση δεδομένων επίδοσης από Raspberry Pi και Jetson Nano κατά τη διάρκεια εκτέλεσης αλγορίθμων νευρωνικών δικτύων.

Στην εργασία αυτή θα πρέπει να εκτελέσετε (inference) διάφορους αλγόριθμους νευρωνικών δικτύων βαθιάς μάθησης για συγκεκριμένα δεδομένα και κάτω από συγκεκριμένες συνθήκες προκειμένου να καταγράψετε και εν συνεχεία να αναλύσετε τις τιμές επίδοσης διαφόρων συστημάτων (ειδικά Raspberry PI και Jetson Nano). Αναζήτηση συσχετισμών μεταξύ χαρακτηριστικών μοντέλων νευρωνικών δικτύων βαθιάς μάθησης και επίδοσης συστήματος κατά την εκτέλεση του μοντέλου.

Generate Adversarial Examples with GANs

Generative Models such as Generative Adversarial Networks (GANs) can be leveraged to generate more realistic adversarial examples.This task focuses on exploring the potential of GANs to create adversarial examples capable of deceiving machine learning models into misclassifying the provided samples.

Reference: Xiao, C. et al. (2018) ‘Generating adversarial examples with adversarial networks’ , Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, pp. 3905–3911. doi:10.24963/ijcai.2018/543.

Use Generative Adversarial Networks (GANs) to protect classifiers against adversarial attacks.

This task focuses on employing GANs as a defense mechanism against adversarially perturbed samples. A GAN can be trained to model the distribution of clean, unperturbed data. Given a perturbed image, the model finds a close output which does not contain the adversarial changes and, as a result, purifies the attacked sample.

Reference: Dai, X., Liang, K. and Xiao, B. (2024) ‘Advdiff: Generating unrestricted adversarial examples using diffusion models’, Lecture Notes in Computer Science, pp. 93–109. doi:10.1007/978-3-031-72952-2_6.

Generate Adversarial Examples with Diffusion Models

The objective of this task is to utilize Diffusion Models to generate adversarial examples aimed at causing misclassification. Diffusion models can achieve adversarial attacks by incorporating adversarial objectives in their generative process.

Reference: Samangouei, P. (2018) ‘Defense-GAN: Protecting classifiers against adversarial attacks using generative models’ , arXiv preprint, arXiv:1805.06605.

The use of Diffusion Models for Adversarial Purification

In this task, the main purpose is to use Diffusion Models in order to purify adversarial samples. The given adversarial sample is diffused with noise following a forward diffusion process, and then a clean image is recovered following a reverse generative process.

Reference: Nie, W., Guo, B., Huang, Y., Xiao, C., Vahdat, A. and Anandkumar, A. (2022) ‘Diffusion models for adversarial purification’ , arXiv preprint, arXiv:2205.07460.

Comparative study of the convergence speed Vs resilience tradeoff in model aggregation mechanisms in Federated Learning

Caching and Replication Mechanisms for the Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an emerging open standard for integrating external tools, data sources, and reasoning modules with AI models and agents. As MCP-based systems scale, multiple components (clients and servers) may repeatedly request the same data or tool outputs, leading to redundant computation and increased latency. Applying distributed systems concepts such as caching and replication can significantly improve MCP performance, reliability, and scalability. This thesis explores how these mechanisms can be designed and evaluated in the MCP ecosystem.

Reference: None

Design and Implementation of an Event-Driven Extension for the Model Context Protocol (MCP)

The Model Context Protocol (MCP) enables standardized communication between AI models, tools, and data sources. Currently, MCP interactions are largely request–response based — a client must explicitly query a server or tool for data. However, many real-world applications (IoT systems, monitoring dashboards, or multi-agent AI environments) benefit from event-driven communication, where components react automatically to data changes or external triggers. Integrating an event-driven (publish–subscribe) mechanism into MCP would enhance its flexibility, reduce polling overhead, and allow tools to respond in real time to updates from other components.

Reference: None

Coordinated Execution of Multiple Early-Exit AI Models in Resource-Constrained Edge-Cloud Environments

This thesis will investigate the integration of early-exit neural network architectures into collaborative edge - cloud AI systems to enhance computational efficiency and reduce communication overhead. This thesis will explore advanced exiting criteria beyond entropy, such as reinforcement learning-based policies, maximum probability thresholds, and budget-aware mechanisms to improve decision-making under uncertainty. The thesis will quantify the trade-offs between accuracy, latency, and network utilization, comparing selective offloading with traditional non-selective methods. Additionally, it examines the deployment of multiple early-exit models across distributed edge nodes to assess scalability, resource utilization, and potential bottlenecks, ultimately contributing to more efficient, adaptive, and uncertainty-aware AI inference in real-world edge-cloud environments.

Reference: BranchyNet: Fast Inference via Early Exiting from Deep Neural Networks • Resource-aware Deployment of Dynamic DNNs over Multi-tiered Interconnected Systems • EdgeBoost: Confidence Boosting for Resource Constrained Inference via Selective Offloading Split Computing and Early Exiting for Deep Learning Applications: Survey and Research Challenges