Skip to content

Papers I find interesting, mostly focused around the intersection of security, privacy, and ML.

Notifications You must be signed in to change notification settings

andli28/Research-Papers

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

36 Commits
 
 

Repository files navigation

Research-Papers

These are the papers I find interesting, mostly focused around the intersection of security, privacy, and ML. I may also list papers relating to the fundamentals of ML/FL infrastructure, or topics involving AI alignment and fairness. There also might be non-papers in here! I am including whatever helps me grasp the concepts the easiest.

See OpenMined for a brief overview of the types of FL.

This list will be organized by topic and attack model (if applicable).

Table of Contents

Privacy

Defenses

  • IBM (Cloud'22): DeTrust-FL: Privacy-Preserving Federated Learning in Decentralized Trust Setting PDF

Security

Attacks

Model Poisoning

  • (ICML'19): Analyzing Federated Learning through an Adversarial Lens PDF Github
    • Attack Model: "Single, non-colluding malicious agent where the adversarial objective is to cause the model to mis-classify a set of chosen inputs with high confidence."

Defenses

Model Poisoning

  • Federated Learning based on Defending Against Data Poisoning Attacks in IoT PDF

    • Attack Model: "A group of p<n/2 malicious label-flipping poisoning attackers, where n is the total amount of participants’ clients."
  • (NeurIPS'21): FL-WBC: Enhancing Robustness against Model Poisoning Attacks in Federated Learning from a Client Perspective PDF Github

    • Attack Model: "Clients mitigate model poisoning attacks that have already polluted the global model"

Vertical FL

  • Vertical Federated Learning: Challenges, Methodologies and Experiments PDF

FL Optimization

  • Oort: Efficient Federated Learning via Guided Participant SelectionPDF | OSDI 21 🎓
  • (ICML'22): Neural Tangent Kernel Empowered Federated Learning PDF
    • Reduces communication rounds, addresses statistical heterogeneity by transmitting update data that is more expressive than simple model weights/gradients
  • Fed-SNN: Federated Learning with Spiking Neural Networks PDF Github
    • Optimizes for energy efficiency
  • Swan: A Neural Engine for Efficient DNN Training on Smartphone SoCs PDF
  • (ICLR 2021): Federated Learning via Posterior Averaging: A New Perspective and Practical Algorithms PDF Github

FL Systems from big tech companies

Paper

Cross-device

  • Apple: Federated Evaluation and Tuning for On-Device Personalization: System Design & Applications | PDF, PDF
  • Google: Towards Federated Learning at Scale: System Design | MLSys21, Github🎓
  • Meta: Papaya: Practical, Private, and Scalable Federated Learning | MLSys22 🎓

Data Center Architecture

  • Yarn: PDF
  • Omega: PDF
  • Tiresias: A GPU Cluster Manager for Distributed Deep Learning | PDF
  • Leap: Effectively Prefetching Remote Memory | PDF, Github (USENIX'20)🎓
    • Two tricks: Prefetching pages wherever possible
    • Using more efficient data paths that allow them to discard the operating system’s irrelevant disk-access features.

Surveys

  • A survey on security and privacy of federated learning URL
  • Survey on Federated Learning Threats: concepts, taxonomy on attacks and defences, experimental study and challenges PDF

LLMs

  • In AI, is bigger always better? Nature

  • Voyager, An Open-Ended Embodied Agent with Large Language Models Website

    • Vector Database of skills (GPT-4 Generated Code). Keys are descriptions, while the Value is the code of "skills"
  • MemGPT: Towards LLMs as Operating Systems PDF

    • LLMs are constrained by limited context windows, hindering their utility in tasks like extended conversations and document analysis
    • MemGPT manages different memory tiers to provide the appearance of large memory resources through data movement between fast and slow memory (similar to traditional OS virtual context management)
  • Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents arxiv

    • LLMs roleplay as doctors, nurses, patients
    • "After treating around ten thousand patients (real-world doctors may take over two years), the evolved doctor agent achieves a state-of-the-art accuracy of 93.06% on a subset of the MedQA dataset that covers major respiratory diseases."
  • (Perhaps) Beyond Human Translation: Harnessing Multi-Agent Collaboration for Translating Ultra-Long Literary Texts arxiv

  • Titans: Learning to Memorize at Test Time arxiv

    • "We present a new neural long-term memory module that learns to memorize historical context and helps attention to attend to the current context while utilizing long past information"
    • Scales better than transformers for long context windows, maintains high accuracy
  • Fairness through Difference Awareness: Measuring Desired Group Discrimination in LLMs arxiv

    • "Our primary call to action in this work is to bring attention to the important notion of difference awareness."
  • The Big LLM Architecture Comparison (July 2025) Substack

MLSys

  • Hidden Technical Debt in Machine Learning Systems [NeurIPS PDF](https://proceedings.neurips.cc/paper/2015/file/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf
  • Autellix: An Efficient Serving Engine for LLM Agents as General Programs PDF
    • Autellix's approach is to prioritizing calls based on total execution time. They introduce two non-clairvoyant scheduling algorithms that assume no prior workload knowledge of programs.
    • PLAS (Program-Level Attained Service) is for single-threaded programs and ATLAS (Adaptive Thread-Level Attained Service) is for multi-threaded programs represented as general, dynamic DAGs.
    • PLAS prioritizes LLM calls based on the current cumulative service, or execution times, of their source program. ATLAS generalizes that to the maximum cumulative service time across all threads in the same program.
    • The goal is to minimize waiting and enhance performance

Other FL paper lists

Other Lists

About

Papers I find interesting, mostly focused around the intersection of security, privacy, and ML.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published