The Center for Digital Trust's Showcase

lightarti-rest

A Rust library to perform REST requests over the Tor network for mobile applications.

This library builds upon the arti-library. It wraps the arti-library for easy use by mobile applications. Instead of downloading the whole dictionary of the Tor-nodes every hour, a prepared dictionary is downloaded once per week. This reduces bandwidth requirements, as the whole dictionary is around 3MB. The prepared dictionary only weighs 200kB.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

Disco

Decentralized Collaborative Machine Learning

A collection of objects and routines to help you develop your own distributed machine learning algorithm. It has two modes, federated, where a server is keeping track of the model, and decentralized, where each participant has their own local model. You can train your model directly in the browser, allowing it to run on a variety of hardware. Many extensions are in the pipeline, such as adding Byzantine resistance or training without revealing the model. Compared to other solutions, Disco is bridging the gap between people creating machine learning models and people having relevant data. By running without installation nor complex configuration, anyone can help. It's also a breeding ground for new machine learning technologies.

Professor: Martin Jaggi

Lab: Machine Learning and Optimization Laboratory

DEDIS Ledger Architecture

Modular abstractions to perform Distributed Ledger (DL) operations

With Dela you can: learn the architecture of a distributed ledger, run your blockchain / distributed ledger, implement and test your new idea that will revolutionize the blockchain world by adding your new module’s implementation to the Dela ecosystem.

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

d-voting

A decentralized, fully anonymous voting system for EPFL-internal elections.

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

Garfield

System support for byzantine machine learning

Garfield is a library to build Byzantine machine learning (ML) applications on top of popular frameworks such as TensorFlow and PyTorch. We show how to use Garfield to build different architectures for ML applications like single server, multiple workers (SSMW), multiple servers, multiple workers (MSMW), and fully decentralized architecture.

Professor: Rachid Guerraoui

Lab: Distributed Computing Lab

AT2

Asset Transfer System using lightweight broadcast-based primitives instead of consensus

We replace consensus with a lightweight building block -- a broadcast-based primitive -- to obtain a full-fledged asset transfer system (cryptocurrency). This system is more efficient and more scalable compared with Bitcoin and consensus-based solutions. AT2 is the first system to bypass consensus algorithms for implementing asset transfers.

Professor: Rachid Guerraoui

Lab: Distributed Computing Lab

Secure and Privacy-Preserving Credentials for E-ID

Research techniques to improve privacy when using electronic identities

Today's identity security faces challenges like misuse and tracking. Our goal is to enable secure, anonymous, unlinkable E-ID interactions by researching novel cryptographic algorithms. This boosts user trust, creates new business opportunities, and cuts financial losses after data breaches.

Professor: Jean-Pierre Hubaux

Lab: Center For Digital Trust

Combattre le phishing: quelles innovations apporter ?

Combating phishing through vulnerability analysis, detection tools, and best practice guidelines.

This project addresses phishing threats through the analysis of existing countermeasures, identification of vulnerabilities, and development of innovative solutions including best practice guidelines and tools for email configuration testing and malicious link detection. The initiative enhances organizational and societal resilience against phishing attacks through knowledge dissemination and fosters collaboration among stakeholders.

Professor: Jean-Pierre Hubaux

Lab: Center For Digital Trust

Sécurisation de Digiinov Hmi en adéquation avec le Cyber Resilience Act (CRA)

Securing industrial HMI software for EU Cyber-Resilience-Act compliance

This project integrates state-of-the-art software security measures into Digiinov HMI, a platform bridging Operational Technology (OT) and Information Technology (IT) for industrial Human-Machine Interfaces. The collaboration with HEIG-VD and EPFL focuses on achieving compliance with the EU Cyber-Resilience-Act (CRA) requirements before its 2027 enforcement. The security enhancements target the software development lifecycle and maintenance processes for embedded industrial control systems.

Professor: Jean-Pierre Hubaux

Lab: Center For Digital Trust

Klint

Automated verification of network function binaries

Professor: George Candea

Lab: Dependable Systems Lab

Decentralized Privacy-Preserving Proximity Tracing

A privacy preserving exposure notification protocol to inform people about possible exposure to the COVID-19 virus

The goal of the DP3T protocol is to simplify and accelerate the process of identifying people who have been in contact with a positively CoVid-19 tested person, thus providing a technological foundation to help slow the spread of the SARS-CoV-2 virus. The system aims to minimise privacy and security risks for individuals and communities and guarantee the highest level of data protection.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

MLBench

Benchmarking of distributed ML

Framework for distributed machine learning. Its purpose is to improve transparency, reproducibility, robustness, and to provide fair performance measures as well as reference implementations, helping adoption of distributed machine learning methods both in industry and in the academic community. Besides algorithm comparison, a main use case is to help the selection of hardware (CPU, GPU) used to run AI applications, as well as how to connect it into a cluster to get a good cost/performance tradeoff.

Professor: Martin Jaggi

Lab: Machine Learning and Optimization Laboratory

OmniLedger

Ethereum-like blockchain implementation based on skipchains

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

Free GAEN alternative

LibreCHovid - Alternative to GAEN for contact tracing applications

Apple and Google developed the Google-Apple Exposure Notification system in order to provide a uniform API for contact-tracing apps while preserving user privacy. On Android, GAEN is integrated in the Google Play services, which is proprietary. As a consequence, users wanting to keep a free system, as well as phone manufacturers such as Huawei who don't have Google Play services cannot use these applications. This project aims to offer an alternative framework, based on open-source code, which requires only minimal changes to the apps.

Professor: Mathias Payer

Lab: HexHive Group

Lattigo

A lattice-based multiparty homomorphic encryption library in Go

Lattigo is a Go module for Multiparty Homomorphic Encryption, that implements Ring-Learning-With-Errors-based homomorphic encryption primitives. It comprises highly-efficient RNS versions of the BFV and CKKS schemes and their multiparty (N-out-of-N-threshold) counterparts. It also implements a dense-key and sparse-key Bootstrapping procedure for RNS-CKKS. The Lattigo library was originally exclusively developed by the EPFL Laboratory for Data Security until its version 2.4.0. Starting with the release of version 3.0.0, Lattigo is maintained and supported by Tune Insight SA. Lattigo 2.0.0 has been code-reviewed by ELCA in November 2020 and, within the allocated time for the code review, no critical or high-risk issues were found.

Professor: Jean-Pierre Hubaux

Lab: Laboratory for Data Security

Scala.js

Use Scala to develop Web frontends

Scala.js optimizes your Scala code into highly efficient JavaScript. Strong typing guarantees your code is free of silly mistakes. You can use any JavaScript library right from your Scala.js code, either in a statically or dynamically typed way.

Professor: Martin Odersky

Lab: Programming Methods Laboratory

Scala Native

Generate native code from Scala

Native is an optimizing ahead-of-time compiler and lightweight managed runtime designed specifically for Scala. It features low-level primitices, seamless interop with native code and instant startup time.

Professor: Martin Odersky

Lab: Programming Methods Laboratory

Scala 3

Scala 3 release updates and new features

Scala 3 implements many language changes and improvements over Scala 2. It strengthens Scala's foundations, makes the language easier and safer to use, and improves the consistency and expressiveness of language constructs. Key changes include intersection and union types, trait parameters, given instances, opaque type aliases, restricted implicit conversions, and a new metaprogramming model.

Professor: Martin Odersky

Lab: Programming Methods Laboratory

EPFL Logic Synthesis Libraries

Collection of modular libraries for the development of logic synthesis applications

Professor: Giovanni De Micheli

Lab: Integrated Systems Laboratory

EPFL Combinational Benchmark Suite

Benchmarks for logical gates optimizer

Professor: Giovanni De Micheli

Lab: Integrated Systems Laboratory

Stainless

Verification framework for a subset of the Scala programming language

Stainless is a tool for verifying Scala programs developed by the LARA. It can verify that your program is correct for all inputs, it can report inputs for which your program fails when they exist, and it can prove that functions do not loop.

Professor: Viktor Kunčak

Lab: Lab for Automated Reasoning and Analysis

Inox

Constraint solver for Stainless

Inox can be used to prove constraints on methods in a program, to attest the correctness of these parts.

Professor: Viktor Kunčak

Lab: Lab for Automated Reasoning and Analysis

LISA

A Proof Assistant based on first order logic, sequent calculus and set theory.

Professor: Viktor Kunčak

Lab: Lab for Automated Reasoning and Analysis

Scallion

A library for writing parsers and pretty printers for LL(1) languages in Scala.

Professor: Viktor Kunčak

Lab: Lab for Automated Reasoning and Analysis

tinyNF

A simpler and faster NIC driver model for network functions

By merely replacing the state-of-the-art driver with our driver, formal verification of the entire software stack completes in 7x less time, while the verified functions’ throughput improves by 160%.

Professor: George Candea

Lab: Dependable Systems Lab

S²E

Platform for writing tools that analyze the properties and behavior of software systems

S²E is a platform for writing tools that analyze the properties and behavior of software systems. S²E comes as a modular library that gives virtual machines symbolic execution and program analysis capabilities.

Professor: George Candea

Lab: Dependable Systems Lab

SVSHI

Verification framework for smart infrastructures

Professor: George Candea

Lab: Dependable Systems Lab

PIX

Show performance of network functions

Professor: George Candea

Lab: Dependable Systems Lab

eBPF-SE

eBPF symbolic execution tool

eBPF-SE is a tool that symbolically executes eBPF programs written for the Linux kernel. It is based on the KLEE symbolic execution engine and uses models (stubs) for the Linux kernel's libbpf API. This allows exploring all program paths without path explosion from the real API implementation. eBPF-SE was previously part of PIX, a tool that extracts performance interfaces from network function code.

Professor: George Candea

Lab: Dependable Systems Lab

Mozuma

Library of ML standardised models

Mozuma is a library containing a collection of machine learning models with standardised interface to run inference, train and manage model state files.

Professor: Karl Aberer

Lab: Distributed Information Systems Laboratory

ophcrack

Windows password cracker

Ophcrack is a Windows password cracker based on rainbow tables. It is a very efficient implementation of rainbow tables done by the inventors of the method. It comes with a Graphical User Interface and runs on multiple platforms

Professor: Serge Vaudenay

Lab: Security and Cryptography Laboratory

TransFool

Adversarial attack against neural machine translation models

Deep neural networks have been shown to be vulnerable to small perturbations of their inputs. In this paper, we investigate the vulnerability of Neural Machine Translation (NMT) models to these attacks and propose a new attack algorithm called TransFool. TransFool can severely degrade the translation quality for different translation tasks and NMT architectures. Moreover, we show that TransFool is transferable to unknown target models. Finally, based on automatic and human evaluations, TransFool leads to improvement in performance compared to the existing attacks. Thus, TransFool permits us to better characterize the vulnerability of NMT models and outlines the necessity to design strong defense mechanisms and more robust NMT systems for real-life applications.

Professor: Pascal Frossard

Lab: Signal Processing Laboratory

Adversarial Classification aTtack

Classification-Guided Adversarial Attack against NMT

Neural Machine Translation (NMT) models have been shown to be vulnerable to adversarial attacks, wherein carefully crafted perturbations of the input can mislead the target model. In this paper, we introduce ACT, a novel adversarial attack framework against NMT systems guided by a classifier. In our attack, the adversary aims to craft meaning-preserving adversarial examples whose translations in the target language by the NMT model belong to a different class (such as sentiment) than the original translations. Unlike previous attacks, our new approach has a more substantial effect on the translation by altering the overall meaning, which then leads to a different class determined by a classifier. Our attack is considerably more successful in altering the class of the output translation and has more effect on the translation. This new paradigm can reveal the vulnerabilities of NMT systems by focusing on the class of translation rather than the mere translation quality as studied traditionally.

Professor: Pascal Frossard

Lab: Signal Processing Laboratory

Synthetic data privacy evaluation

Privacy evaluation framework for synthetic data publishing

The framework implemented in this library measures the privacy gain of publishing a synthetic dataset in place of the raw data with respect to a specific privacy concern. Each concern is modelled as a privacy adversary that targets an individual record and aims to infer a secret about this record. The library includes implementations of two new privacy attacks on the output of a generative model. To evaluate privacy gain, the framework is instantiated under the chosen threat model and outputs an estimate about how much publishing the synthetic data instead of the raw data reduces the privacy loss of a chosen target record under this threat model.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

Zero-Knowledge Swiss Knife

Library for prototyping composable zero-knowledge proofs in the discrete-log setting

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

Petrelic

Wrapper around RELIC pairings

petrelic is a Python wrapper around RELIC. It provides a simple python interface to the BLS-381 pairing and RELIC's big number class.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

SSCred

Library to facilitate anonymous authentication

SSCred provides the following primitives: 1) Anonymous credential light (ACL) 2) Abe's Blind signature 3) Blinded Pedersen commitment

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

mia

Library for running membership inference attacks (MIA) against machine learning models

These are attacks against privacy of the training data. In MIA, an attacker tries to guess whether a given example was used during training of a target model or not, only by querying the model. See more in the paper by Shokri et al. Currently, you can use the library to evaluate the robustness of your Keras or PyTorch models to MIA.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

VOTEAGAIN

Scalable coercion-resistant voting system

The strongest threat model for voting systems considerscoercion resistance: protection against coercers that force voters to modify their votes, or to abstain. Existing remote voting systems either do not provide this property; require an ex-pensive tallying phase; or burden users with the need to store cryptographic key material and with the responsibility to deceive their coercers. VoteAgain is a scalable voting scheme that relies on the revoting paradigm to provide coercion resistance.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

DatashareNetwork

Decentralized privacy-preserving search engine for investigative journalists

Investigative journalists collect large numbers of digital documents during their investigations. These documents can greatly benefit other journalists’ work. However, many of these documents contain sensitive information. Hence, possessing such documents can endanger reporters, their stories, and their sources. Consequently, many documents are used only for single, local, investigations. DatashareNetwork is a decentralized and privacy-preserving search system that enables journalists world-wide to find documents via a dedicated network of peers.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

WebGraph

Capturing Advertising and Tracking Information Flows for Robust Blocking

ML-based ad and tracker blocker that detects ads and trackers based on their action rather than their content.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

DisparateVulnerability

Disparate Vulnerability to Membership Inference Attacks

A membership inference attack (MIA) against a machine-learning model enables an attacker to determine whether a given data record was part of the model's training data or not. DisparateVulnerability helps in measuring how much a given model can be inverted.

Professor: Carmela Troncoso

Lab: Security and Privacy Engineering Laboratory

PowerSGD

Practical low-rank gradient compression for distributed optimization

New low-rank gradient compressor based on power iteration that can i) compress gradients rapidly, ii) efficiently aggregate the compressed gradients using all-reduce, and iii) achieve test performance on par with SGD. The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD with an optimized communication backend. We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets.

Professor: Martin Jaggi

Lab: Machine Learning and Optimization Laboratory

Distributed Homomorphic Anomaly Detection

Proof-of-concept ML algorithms to do anomaly detection compatible with distributed, encrypted algorithms.

The goal of this project is prove that state-of-the-art anomaly detection algorithms that already achieve over 99% accuracy can be run in a distributed and encrypted fashion. In other words, prove that multiple parties can collectively train a good anomaly detecting model without revealing their data to each other. This has been achieved through transfer learning into a Multi-layer Perceptron (MLP) model which achieved comparable results (95% compared to 99% False-Positive Rate accuracy).

Professor: Martin Jaggi

Lab: Machine Learning and Optimization Laboratory

Meditron

Open-source medical language model for clinical decision support

MEDITRON is a pair of open-source large language models (LLMs) with 7 and 70 billion parameters, tailored to the medical domain. It was trained on carefully curated medical data sources, including peer-reviewed literature and clinical practice guidelines. MEDITRON outperforms other open-source models and closed models like GPT-3.5 on medical benchmarks, coming within 5-10% of GPT-4 and Med-PaLM-2.

Professor: Martin Jaggi

Lab: Machine Learning and Optimization Laboratory

Megatron-LLM

Large language model training library

Megatron-LLM enables pre-training and fine-tuning of large language models (LLMs) at scale. It supports architectures like Llama, Llama 2, Code Llama, Falcon, and Mistral. The library allows training of large models (up to 70B parameters) on commodity hardware using tensor, pipeline, and data parallelism. It provides features like grouped-query attention, rotary position embeddings, BF16/FP16 training, and integration with Hugging Face and WandB.

Professor: Martin Jaggi

Lab: Machine Learning and Optimization Laboratory

AIcrowd

Open platform for data science challenges

Professor: Marcel Salathé

Lab: Salathe Lab

Dynamatic

High-level synthesis compiler

Dynamatic is an academic, open-source high-level synthesis compiler that produces synchronous dynamically-scheduled circuits from C/C++ code. It delivers significant performance improvements compared to state-of-the-art commercial tools in specific situations.

Professor: Paolo Ienne

Lab: Processor Architecture Laboratory

CloudSuite

Benchmark suite for cloud services

CloudSuite is a benchmark suite for cloud services. It consists of a number of applications that have been selected based on their popularity in today's datacenters. The benchmarks are based on real-world software stacks and represent real-world setups.

Professor: Babak Falsafi

Lab: Parallel Systems Architecture Lab

QFlex

State-of-the-art modeling tools for the computer architecture community

QFlex is a full-system cycle-accurate computer architecture simulator of multi-node computer systems.

Professor: Babak Falsafi

Lab: Parallel Systems Architecture Lab

JumboTrace

Java program tracer to record execution state

JumboTrace is an experimental tracer for Java programs that records the evolution of program state during execution to build a trace (sequence of events and executed lines). It works by injecting logging calls into the program using a compiler plugin for javac. The logged events are written to a binary file using serialization, and a minimal frontend can display the traces in the console.

Professor: Clément Pit-Claudel

Lab: Systems and Formalisms Lab

RegElk

Linear regex engine for JavaScript in OCaml

This is a linear regular expression engine for a subset of JavaScript regexes, implemented in OCaml. It supports features like nullable quantifiers, capture reset, and lookarounds with linear time and space complexity. The engine is based on an extension of the PikeVM algorithm, described in the paper "Linear Matching of JavaScript Regular Expressions".

Professor: Clément Pit-Claudel

Lab: Systems and Formalisms Lab

Proof-of-Personhood

Proof that you are an actual person while keeping your anonymity

Fake accounts are a big problem on the internet, while not all people want or can identify themselves using state-issued identities. Proof-of-personhood proposes to use personhood parties where people have to physically show up and get identified as real persons. Using cryptographic primitives, it is then possible to prove that you were part of that group, without revealing your individual identity.

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

Kyber

Cryptographic library

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

DRand

Publicly verifiable randomness

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

HBT

Token for conflicts zone

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

APIR

Authenticated Private Information Retrieval

Professor: Bryan Ford

Lab: Decentralized Distributed Systems Laboratory

RetroWrite

Static instrumentation of binaries for fuzzing and sanitation

RetroWrite can analyze the security of closed source binaries by applying an efficient fuzzing technique that usually only works when the source code is available. This makes it possible to trigger more errors than the default binary fuzzers available, while being much faster in finding these errors.

Professor: Mathias Payer

Lab: HexHive Group

Igor

Crash deduplication through root-cause clustering

Fuzzing has emerged as the most effective bug-finding technique. The output of a fuzzer is a set of proof-of-concept (PoC) test cases for all observed “unique” crashes. It costs developers substantial efforts to analyze each crashing test case. This, mostly manual, process has lead to the number of reported crashes out-pacing the number of bug fixes. Igor is a tool for automated crash grouping/deduplication.

Professor: Mathias Payer

Lab: HexHive Group

datAFLow

Data-flow based fuzzing

DatAFLow is a fuzzer built on top of AFL++. However, instead of a control-flow-based feedback mechanism, it uses a data-flow-based feedback mechanism; specifically, data flows based on def-use associations.

Professor: Mathias Payer

Lab: HexHive Group

Evocatio

Evocatio is a bug analyzer built on top of AFL++ and AddressSanitizer. It automatically discovers a bug’s capabilities.

Evocatio automatically discovers a bug’s capabilities: analyzing a crashing test case (i.e., an input exposing a bug) to understand the full extent of how an attacker can exploit a bug. Evocatio leverages a capability-guided fuzzer to efficiently uncover new bug capabilities (rather than only generating a single crashing test case for a given bug, as a traditional greybox fuzzer does).

Professor: Mathias Payer

Lab: HexHive Group

ViDeZZo

Dependency-aware Virtual Device Fuzzing Framework

ViDeZZo is a virtual device fuzzing framework considering both intra- and inter-message dependencies to balance fuzzing scalability and efficiency.

Professor: Mathias Payer

Lab: HexHive Group

E-ID Demo (Issuer & Verifier)

Demo for the swiss E-ID ledger. It includes an issuer and a verifier.

The issuer is a web application that allows to create a new identity and issue a credential. The verifier is a web application that allows to verify the credential.

Professor: Jean-Pierre Hubaux

Lab: Center For Digital Trust

STORM+

Faster PyTorch optimizer

Professor: Volkan Cevher

Lab: Laboratory for Information and Inference Systems

wehey

WeHe𝕐 - Localizing Traffic Differentiation

This project detects traffic differentiation using end-to-end throughput comparisons between an original network trace and a bit-inverted version that hides any packet payload patterns. It then attempts to localize where the differentiation occurs by analyzing measurements from additional paths that converge at the client's ISP. The localization relies on throughput comparison and loss trend correlation between these paths.

Professor: Katerina Argyraki

Lab: Network Architecture Lab

tero project

Studying Internet latency from Gaming Footage

The system downloads gaming footage from the Twitch streaming platform. It extracts the displayed latency measurements using optical character recognition and knowledge of each game's user interface. It maps streamers to locations using natural language processing on their social media profiles. It analyzes the measurements to produce latency distributions per geographical location.

Professor: Katerina Argyraki

Lab: Network Architecture Lab

gem5-X

gem5-based simulator with architectural extensions

Professor: David Atienza

Lab: Embedded Systems Laboratory

3D-ICE

Thermal simulator for stacked integrated circuits

Professor: David Atienza

Lab: Embedded Systems Laboratory

hdtorch

PyTorch-based hyperdimensional computing library

HDTorch is a PyTorch-based library designed for hyperdimensional computing, featuring custom CUDA extensions to accelerate hypervector operations. It is utilized for analyzing HDC benchmark datasets, focusing on accuracy, runtime, and memory usage. The workflow involves initializing basis vectors, encoding features, and aggregating vectors of the same class.

Professor: David Atienza

Lab: Embedded Systems Laboratory

x-heep

Open-source RISC-V system-on-chip for healthcare monitoring

The X-HEEP project is an open-source RISC-V microcontroller designed for continuous healthcare monitoring. It provides a customizable microcontroller unit (MCU) composed of an open-source, industrial-grade RISC-V CPU, common peripherals, and memories, which users can extend with their own accelerators. The project targets both FPGA and ASIC implementations, with support for Xilinx-based FPGAs.

Professor: David Atienza

Lab: Embedded Systems Laboratory

peacok

Persona Commonsense Knowledge for Consistent and Engaging Narratives

PeaCoK is a large-scale persona commonsense knowledge graph, containing approximately 100K human-validated persona facts. It schematizes five dimensions of persona knowledge identified in previous studies of human interactive behaviours, and distils facts in this schema from both existing commonsense knowledge graphs and large-scale pretrained language models. The analysis indicates that PeaCoK contains rich and precise world persona inferences that help downstream systems generate more consistent and engaging narratives.

Professor: Antoine Bosselut

Lab: Natural Language Processing Lab

distilled counterfactual data

Automated generation of high-quality counterfactual data.

DISCO (DIStilled COunterfactual Data) is a method for automatically generating high-quality counterfactual data at scale. It uses a large general language model to generate phrasal perturbations, which are then filtered by a task-specific teacher model to distill high-quality counterfactual data. The method has been applied to natural language inference tasks, demonstrating improved robustness and generalization across distributions.

Professor: Antoine Bosselut

Lab: Natural Language Processing Lab

meditron

Medical language model pretraining

MEDITRON-70B is a large language model (LLM) aimed at democratizing access to medical knowledge. It is an open-source LLM with 7B and 70B parameters, adapted specifically for the medical domain. The model builds on Llama-2 and extends pretraining on a curated medical corpus, including selected PubMed articles and internationally-recognized medical guidelines.

Professor: Antoine Bosselut

Lab: Natural Language Processing Lab

crow

Benchmarking Commonsense Reasoning in Real-World Tasks

CRoW is a manually-curated, multi-task benchmark that evaluates the ability of models to apply commonsense reasoning in the context of six real-world NLP tasks. It is constructed using a multi-stage data collection pipeline that rewrites examples from existing datasets using commonsense-violating perturbations. The study reveals a significant performance gap when NLP systems are evaluated on CRoW compared to humans, indicating that commonsense reasoning is far from being solved in real-world task settings.

Professor: Antoine Bosselut

Lab: Natural Language Processing Lab

FG-NIC

Noise-robust image classification

Image classification has significantly improved using deep learning. This is mainly due to convolutional neural networks (CNNs) that are capable of learning rich feature extractors from large datasets. However, most deep learning classification methods are trained on clean images and are not robust when handling noisy ones, even if a restoration preprocessing step is applied. We propose a method that can be applied on a pretrained classifier. We improve the noisy-image classification (NIC) results by significantly large margins, especially at high noise levels, and come close to the fully retrained approaches.

Professor: Sabine Süsstrunk

Lab: Image and Visual Representation Lab

LELSD

Interactive GAN-based image editing

When generating an image using a GAN, some errors might appear which are obvious to the human observer, but not to the computer. By introducing this user feedback one can refine the generated image to look even more real.

Professor: Sabine Süsstrunk

Lab: Image and Visual Representation Lab

TempSAL

Uncovering Temporal Information for Deep Saliency Prediction

Deep saliency prediction algorithms complement the object recognition features, they typically rely on additional information, such as scene context, semantic relationships, gaze direction, and object dissimilarity. However, none of these models consider the temporal nature of gaze shifts during image observation. We introduce a novel saliency prediction model that learns to output saliency maps in sequential time intervals by exploiting human temporal attention patterns.

Professor: Sabine Süsstrunk

Lab: Image and Visual Representation Lab

DyNCA

Real-Time Dynamic Texture Synthesis Using Neural Cellular Automata

Current Dynamic Texture Synthesis (DyTS) models in the literature can synthesize realistic videos. However, these methods require a slow iterative optimization process to synthesize a single fixed-size short video, and they do not offer any post-training control over the synthesis process. We propose Dynamic Neural Cellular Automata (DyNCA), a framework for real-time and controllable dynamic texture synthesis.

Professor: Sabine Süsstrunk

Lab: Image and Visual Representation Lab

PriCell

Privacy preserving neural network learning for diseased cell classification

PriCell is a Golang library that ensures CellCnn analysis in privacy-preserving, distributed, N-party setting.

Professor: Jean-Pierre Hubaux

Lab: Laboratory for Data Security

CRISP

Offload computation to services while keeping data private

CRISP relies on quantum-resistant lattice-based approximate homomorphic encryption primitives that support flexible polynomial computations on encrypted data without degrading utility. To ensure data integrity, it employs lattice-based commitments and zero-knowledge proofs based on the multi-party-computation-in-the-head paradigm, which enables users to simultaneously convince service providers about the correctness of the encrypted data as well as the authenticity of the underlying plaintext data, using the deployed certification mechanism.

Professor: Jean-Pierre Hubaux

Lab: Laboratory for Data Security

aiFlows

Modular AI collaboration framework

aiFlows is a framework centered around Flows and messages. Flows are independent, self-contained computational building blocks that can complete semantically meaningful units of work. Flows communicate via a standardized message-based interface, enabling modularity, reusability, remote peer-to-peer collaboration, and concurrency.

Professor: Robert West

Lab: Data Science Lab

Transformers CFG

Grammar-constrained text generation with Transformers models

Transformers_cfg is an extension library for the Hugging Face Transformers library that enables grammar-constrained text generation. It provides tools and functionalities to work with context-free grammars (CFGs) for natural language processing tasks involving CFGs. The library supports various Transformer models, including LLaMa, GPT, Bloom, Mistral, and Falcon, and offers features like multilingual grammar support and integration with Text-Generation-WebUI.

Professor: Robert West

Lab: Data Science Lab

Entity Insertion in Wikipedia

Multilingual entity insertion in Wikipedia articles

Proposes a framework for inserting entities into Wikipedia articles across multiple languages. It processes Wikipedia dumps to extract data and train models for entity insertion. The key components are: 1) Data processing pipeline to extract relevant data from Wikipedia dumps. 2) Modeling code for training entity insertion models using a ranking loss or pointwise loss. 3) Benchmarking code to evaluate models against baselines like BM25, EntQA, and GPT language models.

Professor: Robert West

Lab: Data Science Lab

Refiner

Reasoning framework with feedback on intermediate steps

REFINER is an interaction-based framework for natural language reasoning tasks. It has a CRITIC model that provides structured feedback on intermediate reasoning steps, and a GENERATOR model that solves the reasoning task by first generating intermediate steps. The core idea is the interaction between the generator and critic, where the generator's steps are improved via feedback from the critic.

Professor: Robert West

Lab: Data Science Lab

LAMEN

Evaluating language models through negotiation tasks

The project proposes using structured negotiations as a dynamic benchmark for evaluating language model (LM) agents. The negotiation framework consists of a game setting, issues to negotiate, and optional preference weights, allowing for the design of complex games by increasing the number of issues, mixing issue types, and adding non-uniform preferences. The benchmark setup jointly evaluates performance metrics (utility and completion rate) and alignment metrics (faithfulness and instruction-following) in self-play and cross-play settings.

Professor: Robert West

Lab: Data Science Lab

LLM Grounding Analysis

LLM grounding vs. factual recall

The study investigates LLMs using Fakepedia, a dataset presenting contradictions between known facts and new information. Through Masked Grouped Causal Tracing (MGCT), the research deciphers LLMs' grounding mechanisms by contrasting neural activation patterns. Findings help understand the co-functioning of grounding with recall capabilities within LLMs.

Professor: Robert West

Lab: Data Science Lab