Positions: 3

Research Grant (BI)

  • Project PT Smart Retail–RefªC6632206063-00466847-BI|2024/529

    Type of position: Research Grant (BI)
    Duration: 6 months
    Deadline to apply: 2024-05-09
    Description

    The goal of this project is to contribute to the development of advanced security and privacy-preserving technologies for smart-retail stores. Specifically, this project aims to study an emerging class of vulnerabilities named Denial of Wallet (DoW) attacks. In this type of attack, attackers can exploit vulnerabilities in cloud-hosted services that can trigger excessive resource usage, such as external APIs or public storage, resulting in financial damage to smart retail actors that rely on the cloud for hosting their web applications. Considering these attacks, this project has the following main goals: (1) Design and implement DoWGuard, a novel detection tool for DoW vulnerabilities in cloud applications; (2) Extend the existing static analysis techniques to include the ability to reason about economic sinks and interactions with cloud APIs. (3) Define and specify queries within DoWGuard to accurately identify DoW vulnerability patterns. (4) Integrate DoWGuard into the CI/CD toolchain, providing developers with immediate feedback on potential DoW vulnerabilities during the development lifecycle. (5) Evaluate DoWGuard's effectiveness by applying it to real-world cloud applications. The expected outcome of this project includes both a report of the analyzed solutions and a software prototype of the implemented solution.


    Contact email: rh@inesc-id.pt

  • Project ACCELERAT.AI – refªC644865762-00000008-BI|2024/527

    Type of position: Research Grant (BI)
    Duration: 4 months
    Deadline to apply: 2024-05-06
    Description

    A significant number of transformer-based language models specifically tailored for European Portuguese have been recently proposed, such as Albertina, Gervásio or Glória, among others.  These type of models have shown exceptional modelling capabilities of language (understanding and generation) and remarkable performance on a wide range of natural language processing tasks. Concurrently, a similar effort by the research community has resulted in the proposal of several foundation models for speech and audio. These models are trained on extensive amounts of multilingual data to acquire robust representations that capture the intrinsic structure and relationships within speech and audio signals.

    The goal of this project is to delve into the existing Portuguese language model landscape, conducting comparative analyses and assessing their potential utility. The final aim is to integrate these models into an automatic speech recognition system. This system will employ a pre-trained speech encoder, one of the studied pre-trained large language models (utilizing either encoder-decoder or decoder-only configurations), and a neural adaptor module.  

    The work plan includes the following steps:

    1. Research existing foundation models for European Portuguese, both text and speech/audio.

    2. Analyse and compare them in common benchmarks.

    3. Assess their potential utility as a part of an ASR pipeline that integrates both pre-trained speech encoders and LLMs.


    Contact email: rh@inesc-id.pt

Contract