1. Abstract

Problem: Ethiopian smallholder farmers face increasing challenges from climate change, including erratic rainfall and soil degradation. While data-driven agriculture can offer solutions, traditional centralized systems risk exposing sensitive farm data (e.g., precise location, yields), leading to privacy concerns and lack of adoption.

Motivation: There is a critical need for a system that can provide personalized, climate-smart agricultural advice without compromising the privacy of individual farmers. This project aims to bridge the gap between advanced AI capabilities and data sovereignty.

Approach: We developed a Privacy-Preserving Multi-Agent System that utilizes Federated Learning (FL) principles and Differential Privacy (DP). The system consists of five specialized agents orchestrated by LangGraph. It processes sensitive data locally on the "farm" (simulated), aggregates insights using DP noise to mask individual contributions, and generates context-aware recommendations using Retrieval-Augmented Generation (RAG) with real-time climate data.

Key Outcomes: The system successfully generates personalized crop and weather plans while maintaining a high level of privacy. Benchmarks demonstrate that lower privacy budgets (epsilon < 0.5) significantly reduce the risk of membership inference attacks (success rate < 5%) while preserving the utility of the aggregated insights for regional trend analysis.

2. Methodology

2.1 System Architecture

The system is built on a modular multi-agent architecture using LangGraph for orchestration. The workflow mimics a federated learning cycle:

1. Local Analysis: Data stays on the device.
2. Federated Collaboration: Only gradients/insights are shared, not raw data.
3. Planning & Audit: Recommendations are generated and audited for privacy leaks before delivery.

graph TD
    User[User Input] --> Local[Local Data Analyzer]
    Local -->|Anonymized Features| Fed[Federated Collaborator]
    Fed -->|Regional Trends| Planner[Crop/Weather Planner]
    Planner -->|Draft Plan| Auditor[Privacy Auditor]
    Auditor -->|Approved| Synth[Synthesizer]
    Auditor -->|Rejected| Planner
    Synth -->|Final Report| Output[Multilingual Output]

2.2 Agents and Roles

• Local Data Analyzer: Uses gpt-4o to extract features (soil pH, crop type) from natural language input. It applies local differential privacy noise before any data leaves this node.
• Federated Collaborator: Simulates a central server that aggregates updates from multiple "farms" (peers). It uses the Laplace Mechanism to add noise to the aggregate, ensuring that no single farmer's data significantly influences the global model.
• Crop/Weather Planner: Combines local anonymized data with regional trends. It utilizes Tavily Search to fetch real-time climate resilience data (e.g., rainfall patterns, drought-resistant varieties) to ground its recommendations.
• Privacy Auditor: Acts as a guardrail, checking the final output for any potential PII leaks or sensitive data exposure before it reaches the user.
• Synthesizer: Formats the final recommendation into a clear, actionable report and translates it into English, Amharic, and Afaan Oromoo.

2.3 Technologies Used

• LangGraph: For stateful multi-agent orchestration.
• LangChain: For LLM interaction and tool usage.
• OpenAI GPT-4o: The core reasoning engine.
• Tavily API: For real-time web search (RAG).
• NumPy: For numerical simulations and differential privacy calculations.

3. Results

3.1 Privacy Benchmarks

We evaluated the system's resilience against Membership Inference Attacks, where an adversary attempts to determine if a specific user's data was included in the aggregated result.

Test Setup:

• Attack Simulation: We simulated an adversary guessing membership based on the difference between the true population mean and the DP-noised mean.
• Metric: Attack Success Rate (lower is better).

Findings:

Note: A success rate of ~0.0 implies the attack is no better than random guessing in this simplified model.

3.2 Qualitative Outcomes

• Personalization: The system correctly identifies specific needs (e.g., acidic soil treatment for Teff) based on local analysis.
• Safety: The Privacy Auditor successfully flags and blocks outputs that inadvertently contain raw location coordinates or PII.
• Accessibility: The multilingual synthesis ensures that advice is accessible to farmers in their native languages.

3.3 Conclusion

The Privacy-Preserving Multi-Agent Climate-Resilient Farming Advisor demonstrates that it is possible to leverage the power of modern AI and collaborative learning without compromising the digital sovereignty of smallholder farmers. By embedding privacy by design via Federated Learning and Differential Privacy, we pave the way for more trustworthy and scalable digital agriculture solutions.