Master’s & Research projects for EPFL students
The following Master’s and Research projects are available to EPFL students, with options for both individual and collaborative work.
To apply, send an email from your EPFL address to Prof. Andrea Cavallaro with the subject line “MINTS project: [Your chosen project title]”. Please attach your current CV and most recent academic transcript, and include your motivation paragraph directly in the body of the email.
Contrastive latent alignment for multimodal reasoning
Brief description: Sensory inputs containing conflicting signals, such as a visual scene hallucination contradicting audio cues or text context, generate corrupted latent representations. Unlike text chain-of-thought where humans can inspect intermediate reasoning traces to detect visual-textual misalignment, continuous hidden states are (currently) uninterpretable. The student(s) will explore a contrastive latent alignment process that enforces cross-modal and preference alignment within continuous latent reasoning paths by integrating, for example, cross-modal contrastive trajectory optimization with direct preference alignment over continuous latent manifolds.
Locating and ablating circuit-level private attribute sub-networks
Brief description: AI models ingest vast corpora containing sensitive or personally identifiable information, or PII. Traditional machine unlearning methods may reduce model utility and/or induce catastrophic forgetting. Instead of retraining or fine-tuning model parameters, the student(s) will explore activation-space dictionary learning to map the circuits responsible for encoding specific private attributes, ablate or orthogonalize these latent feature directions, and (ideally) guarantee PII elimination while preserving zero-shot task performance.
Continuous latent flow matching for multimodal spatial reasoning
Brief description: To plan physical interactions, Vision-Language-Action models mainly rely on step-by-step auto-regressive chain-of-thought generation in natural language or discrete token space. While recent continuous latent reasoning methods allow models to ‘think’ in hidden representation spaces, existing frameworks generally need multi-step latent iterations. The student(s) will explore a single-step continuous latent flow architectures that replaces iterative chain-of-thought token generation with a single-step vector-field flow mapping directly from multimodal inputs (video, audio, natural language instructions) for latent action-reasoning representations. The goal is to achieve high-accuracy spatial problem-solving in a single forward pass.
Alignment-based conformance checking
Brief description: Organizations increasingly automate enterprise workflows using multi-agent systems. Ensuring real-time adherence to regulatory frameworks, such as the EU AI Act or General Data Protection Regulation (GDPR), has become an important operational bottleneck. The student(s) will explore a neuro-symbolic conformance checking framework based on semantic parsing and symbolic temporal logic. The student(s) will develop a verifiable, transparent mechanism to detect and explain compliance risks or failures by translating natural language regulatory articles into linear temporal logic specifications and evaluating them with execution traces.
Visual stream parsing for compliance auditing
Brief description: AI-driven stream analytics proliferate across public and enterprise environments. Ensuring continuous compliance with regulatory instruments, such as the EU AI Act and General Data Protection Regulation (GDPR), has become an important engineering objective. The student(s) will design a dynamic legal-visual (or legal-audio-visual) concept engine that scales static concept probing to real-time video streaming. By fusing vision-language (or audio-visual-language) transformer streaming backbones with ontology graph mapping, the student(s) will build an end-to-end framework capable of tracking entities, inferring sensitive personal attributes and mapping scene semantics to legal frameworks.
Voice anonymization with preservation of emotional prosody in multilingual settings
Brief description: Voice anonymization methods may modify the acoustic voiceprint through pitch shifting, formant warping or latent space transformations to mask physical biometrics. These methods often degrade emotional expressiveness or impact the performance of automatic speech recognition. The student(s) will design, implement and evaluate an AI-based acoustic transformation framework that masks speaker identity while preserving emotional prosody, semantic meaning and acoustic intelligibility. The primary focus is French/Italian or German, including Swiss German variants. Evaluation with human participants and real-world datasets will be carried out in collaboration with Radio Télévision Suisse (RTS).
Linguistic and stylometric style transfer for long-form speech anonymization
Brief description: In long-form recordings, an adversary can re-identify a speaker by analyzing persistent linguistic markers such as accent, dialect variations, syntax, vocabulary and unique phrasing habits. Traditional anonymization only redacts explicit Personally Identifiable Information, or PII, like names and locations and generally do not account for speaking styles. The student(s) will leverage sequence-to-sequence style-transfer models and design an AI framework to translate/obfuscate individual speaking styles, syntax and dialectal markers. The primary focus is French/Italian or German, including Swiss German variants. Evaluation with human participants and real-world datasets will be carried out in collaboration with Radio Télévision Suisse (RTS).
Privacy guarantees for speech anonymization
Brief description: This project covers AI signal processing, deep generative modeling and privacy theory. Inference attacks on voice recordings exploit the correlation of acoustic characteristics, emotional prosody and linguistic stylometrics. Anonymization benchmarks rely primarily on empirical measures such as equal error rate. The student(s) will build a framework to explicitly bound and minimize mutual information across modalities while providing information-theoretic privacy guarantees against adaptive or semi-informed adversaries who leverage AI models trained on modified speech. Evaluation with human participants and real-world datasets will be carried out in collaboration with Radio Télévision Suisse (RTS).