RT @openminedorg: AI researchers & dataset owners — The NAIRR Pilot seeks datasets to expand AI education across the U.S. Ideal datasets:….

@openminedorg Final slide, feel free to check the QR codes and get all the needed materials. Do not forget to join the @openminedorg Community. P.S Sorry Valerio for not being able to tag you properly @leriomaggio

@openminedorg #SyftBox @openminedorg .Rather than sharing any data, we can share federated computations through the network. PySyft vs SyftBox: every folder is a datasite. Pysyft has a clear distinction between who is the data provider and the computations

@openminedorg For anyone interested: An open-use case is available below combining SyftBox and FL .@openminedorg

@openminedorg Federated Learning is a genuine ML community. It works smoothly and simply. "Why not bring the data where the model lives!" says Valerio. @openminedorg.Federated averaging by returning averages to clients without letting them look at the data. FL can be combined with other PETs

@openminedorg I like this: "Not all PETs work for all use cases". >> There is no-1-size fits all. @openminedorg #20DaysOfFLChallenge.

@openminedorg K-anonymization vs Differential privacy

@openminedorg Getting into something that I am currently testing and using in my experiments, the so-called #Pysyft by @openminedorg

@openminedorg Getting slowly into a distributed setting. "We move the computation where data lives". Check this paper: Structured Transparency Framework: a systematic approach to govern to create trust between parties. It works using policies in: #input, #model, #output. Amazing!

@openminedorg "The data vs privacy dilemma".> PETs are the techniques to enable us to work with technologies in a privacy-preserving manner. - The so famous "Privacy vs Utility Tradeoff"

@openminedorg #PETs .Privacy-Enhancing Technologies are a joint effort from different #communities. PPML vs PETs: #PPML is a subset of #Pets. Self-notes: Nice to see #SyntheticData in the slide. Really enjoying to see the slides from @openminedorg They make simple definitions and coverage.

@openminedorg Issue #2: Reliable ML models. Even if we think that we are running data locally in a private env it can be under attack e.g. Re-identification (reveal identity), inference (infer sensitive information), membership inference, model inversion attacks.#SharingisCaring

@openminedorg Issue #1: Data availability :.not all the data is available or easily accessible. >> "You cannot just download data". #30DaysOfFLChallenge @openminedorg

#30DaysOfFLCode . Happening now, an ongoing session on introduction to #privacy-preserving #ML by Valerio Maggio from @openminedorg . Registration (access to the link) : .It will be the next one hour. If not able to join, see below my learned lessons :)

@iamtrask 3/ Exploring more repositories & recommendation algorithms to push the boundaries of my knowledge, and understanding. Always open to suggestions or discussions! .#RecommenderSystems.

@iamtrask 2/ Next up: Attempting to reproduce Gustavo Bertoli's work [Link: ! .Hit a snag with the dummy CSV file 📂, but it's all part of the learning process. #FederatedLearning #PETs #AI.

#30DaysOfFLCode .1/ 🚀 Diving deep into privacy-enhancing technologies (PETs) by reviewing some papers & running experiments! .I had my hands on @iamtrask's Netflix FL code. The code is self-explanatory and easy to follow. Full details here: 🖥️.

4️⃣ Outstanding Results: Demonstrates up to 34% HR@10 and 42% NDCG@10 improvements across datasets, all under a reasonable privacy budget (e.g., 𝜖 = 1).

@openminedorg 2️⃣ Semantic Recovery: Introduces a semantic-preserving publishing method to locally perturb user patterns, recovering meta-path semantics while ensuring privacy. 3️⃣ HGNN Design: Develops a model that performs node and semantic-level aggregations to capture recovered HIN semantics.

@openminedorg The authors propose FedHGNN, a federated heterogeneous graph neural network framework for privacy-preserving recommendations using distributed HINs. Contributions:.1️⃣ Privacy Definition: Protects user-item interactions & high-order patterns using differential privacy.