Internship Research Portfolio · INFILTRATE

Aseel KHATIB

Designing a negotiation mechanism for open, federated, service-based learning systems — Months 1–3
Mission

Internship Objective

Core Goal: Design and implement a negotiation mechanism for open, federated, service-based learning systems — enabling autonomous participants to agree on the terms of their collaboration (data contribution, model access, reward sharing, trust requirements) without a central authority imposing participation.

Scientific Context: Federated Learning assumes participants cooperate. In reality, participants are rational agents with private costs, competing interests, and heterogeneous data quality. Negotiation is the mechanism that transforms a set of self-interested parties into a functioning federation — making it the missing layer between FL theory and open FL practice.

Method: Three phases over three months: (1) literature review on FL, Open FL, Trust, and Negotiation — building the theoretical foundation; (2) comparative study of FL platforms to identify which can host a negotiation layer; (3) design and first implementation of the negotiation protocol on the selected platform.

The key reference paper (Bena et al.) provides the architectural context: a Trust Management System for collaborative, federated, service-based applications. The negotiation mechanism is the component this internship contributes to that architecture — answering the question of how parties agree to join, what they commit to, and what they receive in return.

Research Questions
  1. RQ1 — Agreement: What are the dimensions of a participation agreement in open FL — what must be negotiated (contribution level, reward, privacy budget, trust threshold) and by whom?
  2. RQ2 — Protocol: What negotiation protocol (bilateral, multilateral, auction-based, bargaining-based) best fits the open FL context, balancing fairness, efficiency, and computational cost?
  3. RQ3 — Trust-Negotiation Coupling: How should trust scores (from the TMS of Bena et al.) influence negotiation outcomes — and how should negotiated commitments update trust scores over time?
  4. RQ4 — Platform Support: Which FL platform best supports the implementation of a negotiation layer, and what extensions are required to host the negotiation protocol alongside FL training?
Internship Objective

Evaluation Dimensions

DimensionConcrete Test for FedNegotFlowerTFF
Ease of Use Time from install to first running FL experiment on the use case. Number of lines of boilerplate required.
Technical Characteristics Sync vs async rounds; client selection mechanism; how model weights are serialised and transmitted.
Flexibility of Use Can aggregate_fit() be replaced with trust-weighted aggregation without forking the library?
Dynamicity Possibility Can a new client join between rounds? Can we drop a client mid-training without crashing?
Extendability How many lines / files must be added to host the full FedNegot negotiation layer (pre-round + post-round hooks)?

Proposed Architectures

FedNegot — Four Platform Deployment Architectures