Patwardhan, Narendra Prakash (2025) Building Secure Embodied Agents for Personalized Environments. [Tesi di dottorato]
Questa è la versione più aggiornata di questo documento.
![[thumbnail of Thesis_Building_Secure_Embodied_Agents.pdf]](https://www.fedoa.unina.it/style/images/fileicons/other.png "Thesis_Building_Secure_Embodied_Agents.pdf") |
Documento PDF
Thesis_Building_Secure_Embodied_Agents.pdf Download (8MB) |
| Tipologia del documento: | Tesi di dottorato |
|---|---|
| Lingua: | English |
| Titolo: | Building Secure Embodied Agents for Personalized Environments |
| Autori: | Autore Email Patwardhan, Narendra Prakash narendraprakash.patwardhan@unina.it |
| Data: | 9 Dicembre 2025 |
| Numero di pagine: | 288 |
| Istituzione: | Università degli Studi di Napoli Federico II |
| Ciclo di dottorato: | 38 |
| Coordinatore del Corso di dottorato: | nome email Russo, Stefano stefano.russo@unina.it |
| Tutor: | nome email Sansone, Carlo [non definito] |
| Data: | 9 Dicembre 2025 |
| Numero di pagine: | 288 |
| Parole chiave: | Embodied Agents, Capability-Based Security, Parameter-Efficient Fine-Tuning, Edge Computing, Ambient Intelligence |
| Settori scientifico-disciplinari del MIUR: | Area 09 - Ingegneria industriale e dell'informazione > ING-INF/05 - Sistemi di elaborazione delle informazioni |
| Informazioni aggiuntive: | 38° ciclo di dottorato |
| Depositato il: | 10 Dic 2025 22:36 |
| Ultima modifica: | 12 Ago 2026 05:39 |
| URI: | https://www.fedoa.unina.it/id/eprint/17105 |
Available Versions of this Item
-
La villa di Don Luise di Toledo a Napoli. (deposited 23 Nov 2009 14:04)
- Building Secure Embodied Agents for Personalized Environments. (deposited 10 Dic 2025 22:36) [Attualmente visualizzato]
Abstract
Ambient Intelligence envisions environments that anticipate human needs through embedded computational systems. While Large Language Models demonstrate suitable reasoning capabilities, their deployment as embodied agents faces architectural constraints: unsandboxed execution of model-generated code, coupling to specific runtime environments, and personalization requiring either intensive computation or cloud transmission. This dissertation presents an integrated architecture combining transparent foundation models, formally verified execution substrates, and gradient-free adaptation. The contribution comprises three components: language models at 0.6B, 8B, and 14B parameters trained through staged protocols with modular architectures enabling efficient specialization; an execution framework implementing capability-based security via WebAssembly with formal operational semantics, achieving provable isolation while reducing privileged operations by 95%; and an adaptation mechanism applying local learning rules to low-rank projections, enabling continuous personalization with 12.5× sample efficiency over supervised methods while maintaining 99.6% knowledge retention. A Smart Couch system validates the architecture, deploying a 0.6B model on embedded hardware for multimodal health monitoring and comfort optimization. The system achieves sub-200ms inference latency and 96.3% task completion accuracy while processing all data locally, demonstrating that formal security, efficient adaptation, and sophisticated reasoning coexist in resource-constrained embodied agents.
Downloads
Downloads per month over past year
Actions (login required)
 |
Modifica documento |
