ASILO — Adaptive Swarm-Inspired Local Optimisation (Fully Decentralized FL)

Prototype + experiments for fully decentralized federated learning (no central server) on non-IID and heterogeneous clients, with a focus on communication efficiency, robustness, and fairness.

Repository (thesis code): https://github.com/Saurabh97/Asilo_1

---

What this project is about

Fully decentralized FL is hard because:

• Non-IID data: each client has different label distributions and behavior.
• Heterogeneous clients: devices differ in compute budget (local epochs/steps), bandwidth, and reliability.
• Communication constraints: limited fan-out, byte budgets, message drops/buffering.
• Fairness: some clients can dominate the learning while others get ignored.

This repo implements ASILO (a swarm-inspired decentralized method) and compares it against multiple baselines.

---

Key idea of ASILO

ASILO combines:

• Pheromone-guided peer selection (learns which neighbors are high-utility)
• Capability-aware local optimisation (clients train according to their own budgets)
• Event-triggered / lightweight communication (communicate when useful)
• Robust gating + aggregation (accept/buffer/drop updates based on similarity/budget/robust rules)

The goal is to improve the accuracy–communication tradeoff while maintaining stability and fairness across clients.

---

Implemented methods

• ASILO — swarm-inspired, decentralized, capability-aware, event-triggered communication
• FedAvg — classic FL baseline adapted to this framework
• DeceFL — peer-based decentralized update exchange
• DFedSAM — decentralized training with SAM optimizer support

Note: configs are designed so that local compute budget comes from YAML (e.g., local_round, local_epochs, or steps-per-round depending on trainer).

---

Dataset

Experiments target WESAD (wrist wearable signals) with:

• strong subject-level non-IID
• label imbalance and varying class coverage across subjects
• common setups include binary stress vs non-stress (depending on preprocessing)

---

Training pipeline highlights

To make FL rounds realistic and comparable across heterogeneous clients:

• Round-wise training budget (fixed steps/epochs per round)
• Streaming DataLoader iterator (no full-epoch loop unless desired)
• iterator reinitializes on exhaustion (optional reshuffle)
• optional replay buffer growth for online/streamed data
• gradient clipping + safe evaluation guards

---

Configuration (YAML)

Most experiments are driven by YAML configs (example structure):

round_time_s: 0.1

run_limits:
  max_rounds: 100
  max_seconds: 1200

agents:
  - { id: S2, host: 127.0.0.1, port: 9202, local_round: 5,  model_id: wear }
  - { id: S3, host: 127.0.0.1, port: 9203, local_round: 10, model_id: wear }
  - { id: S4, host: 127.0.0.1, port: 9204, local_round: 5,  model_id: wear }

# method-specific blocks:
asilo:
  k_peers: 3
  trigger: ...
  pheromone: ...
  agg_mode: "robust"