"Collaborative CRM intelligence without sharing customer data"
Author: Nikhil Reddy Donapati · Agentforce AI Specialist · Texas, USA
Bank A: "Our fraud signals would improve Bank B's models — but we can't share customer data"
Hospital X: "Patient risk patterns could help all providers — HIPAA prevents sharing"
Insurer Y: "Cross-company churn patterns are valuable — GDPR forbids it"
Traditional Answer: Train in isolation → poor generalization
Centralized Answer: Pool all data → legally untenable
FedCRM-DP Answer: Federated learning → collaborative AI, no data sharing
TABLE I: Comparative Performance — All Six Strategies
Strategy Accuracy F1-Score AUC-ROC Privacy Guarantee
─────────────────────────────────────────────────────────────────────
Local-Only 0.8140 0.3850 0.7420 None (isolated)
FedAvg 0.8750 0.5210 0.8540 None (gradient leakage)
FedProx 0.8890 0.5840 0.8810 None (gradient leakage)
FedAdam 0.8480 0.5718 0.9338 None (gradient leakage)
FedCRM-DP ε=5 ★ 0.8850 0.5720 0.8760 (ε,δ)-DP + SecAgg ✓
Centralized † 0.9120 0.6480 0.9320 Not deployable (GDPR/HIPAA)
★ Proposed | † Upper bound only
FedCRM-DP retains 97.0% of FedProx F1 with formal privacy guarantees.
vs FedAvg: t=7.73, p=0.0015 | vs FedProx: t=−3.09, p=0.037
FedCRM-DP (ε=5): Acc=0.7840±0.0038, F1=0.6165±0.0022, AUC=0.8504
+1.79pp vs FedAvg (t=9.35, p=0.0112) | +1.74pp vs FedProx (t=9.13, p=0.0118)
FedCRM-DP: Four-Layer Federated Stack
┌───────────────────────────────────────────────────────────────────┐
│ L1 LOCAL ENTERPRISE TRAINING │
│ Raw data NEVER leaves organizational boundaries. │
│ Five demographically partitioned clients (Bank Marketing): │
│ Client A (age<30, n≈7.2k) | Client B (age30-45, n≈17.5k) │
│ Client C (age46-60, n≈12.4k) | Client D (age>60, n≈5.1k) │
│ Client E (mixed, n≈3.0k) │
├───────────────────────────────────────────────────────────────────┤
│ L2 FEDPROX OPTIMIZATION │
│ F_k(w) = f_k(w) + (μ/2)||w − w_t||² (μ=0.01) │
│ Bounds client drift under non-IID demographic distributions │
│ +6.3pp F1 over FedAvg in heterogeneous CRM settings │
├───────────────────────────────────────────────────────────────────┤
│ L3 DIFFERENTIAL PRIVACY (DP-SGD) │
│ Per-sample gradient clipping: C=1.0 │
│ Gaussian noise: N(0, σ²C²I), σ calibrated per ε │
│ (ε,δ)-DP with δ=1e-5, ε∈{1,3,5,10} │
│ Pareto-optimal: ε=5 (97% F1 retention, moderate privacy) │
├───────────────────────────────────────────────────────────────────┤
│ L4 SECURE AGGREGATION │
│ Pairwise masking (Bonawitz et al. [5]) │
│ Server learns ONLY weighted mean — not individual updates │
│ w_{t+1} = Σ_k (n_k/n) · w_k^{t+1} │
│ Masks cancel in aggregate; server is blind to contributions │
└───────────────────────────────────────────────────────────────────┘
| ε | Accuracy | F1-Score | Privacy Level | Recommended For |
|---|---|---|---|---|
| 1 | 0.8440 | 0.4910 | Very Strong | Healthcare, regulated sectors |
| 3 | 0.8760 | 0.5520 | Strong | Financial services, GDPR-sensitive |
| 5 ★ | 0.8850 | 0.5720 | Moderate | Standard CRM (Recommended) |
| 10 | 0.8870 | 0.5800 | Minimal | Internal/low-sensitivity data |
git clone https://github.com/nikhilred2619/FedCRM-DP-Framework.git
cd fedcrm-dp
pip install -r requirements.txt
# REST API
uvicorn api.main:app --reload --port 8000
# Run full evaluation (paper replication)
python3 evaluation/experiment.py
# Streamlit demo
streamlit run dashboard/app.py
# Docker
docker compose up --buildimport requests
response = requests.post("http://localhost:8000/fedcrm/predict", json={
"lead_id": "LEAD-001",
"age": 35, "job_category": "management",
"prior_contacts": 2, "campaign_duration": 250,
"balance": 1500, "previous_outcome": "success",
"epsilon": 5.0, "organization_id": "BANK-A"
})
# → {"prediction":1, "probability":0.724, "epsilon":5.0,
# "privacy_level":"Moderate", "compliance_status":"Compliant",
# "explanation":"FedCRM-DP: Likely to convert (72.4%)..."}| Industry | Use Case | Privacy Concern | ε Setting |
|---|---|---|---|
| Banking | Lead conversion prediction | GDPR | 5 |
| Healthcare | Patient risk stratification | HIPAA | 1-3 |
| Insurance | Churn propensity | CCPA | 3-5 |
| Retail CRM | Customer LTV scoring | GDPR | 5 |
| SaaS | Account expansion prediction | SOC2 | 5 |
| HR Tech | Candidate success prediction | EEOC | 3 |
| Configuration | Accuracy | F1 | Privacy |
|---|---|---|---|
| FedProx (no DP) | 0.8890 | 0.5840 | None |
| + DP (ε=10) | 0.8870 | 0.5800 | Minimal |
| + DP (ε=5) ★ | 0.8850 | 0.5720 | Moderate |
| + DP (ε=3) | 0.8760 | 0.5520 | Strong |
| + DP (ε=1) | 0.8440 | 0.4910 | Very Strong |
| Full FedCRM-DP (ε=5) | 0.8850 | 0.5720 | Moderate+SA |
fedcrm-dp/
├── fedcrm_core.py ← Core data structures, paper results
├── federated_engine/orchestrator.py ← Round orchestration, 100-round training
├── local_training/trainer.py ← FedProx local training + DP
├── aggregation_engine/ ← Secure Aggregation (Bonawitz et al.)
├── differential_privacy/engine.py ← DP-SGD: clipping + Gaussian noise
├── api/main.py ← FastAPI REST API (OpenAPI 3.0)
├── crm_integration/
│ ├── salesforce/connector.py ← Salesforce + Agentforce
│ ├── servicenow/connector.py ← ServiceNow workflows
│ ├── sap/connector.py ← SAP customer ops
│ ├── oracle/connector.py ← Oracle compliance
│ └── dynamics365/connector.py ← MS Dynamics 365
├── evaluation/experiment.py ← Full paper replication
├── dashboard/app.py ← Streamlit interactive demo
├── tests/test_fedcrm.py ← 24/24 tests passing
├── results/fedcrm_results.json ← Experimental results
├── Dockerfile + docker-compose.yml
└── requirements.txt
@article{donapati2025fedcrmdp,
title = {FedCRM-DP: A Privacy-Preserving Federated Learning Framework
for Secure Cross-Organization CRM Intelligence},
author = {Donapati, Nikhil Reddy},
year = {2025},
note = {Research manuscript},
url = {https://github.com/nikhildonapati/fedcrm-dp}
}Nikhil Reddy Donapati · Texas, USA
Agentforce AI Specialist & Senior Salesforce Developer