Search papers, labs, and topics across Lattice.
This study evaluates the privacy risks associated with federated learning (FL) in the context of radiology reports by quantifying the extent of information leakage through gradient inversion techniques. By comparing three different tokenizers鈥擥PT-2, RadBERT, and LLaMA-2鈥攁cross multiple datasets and batch sizes, the researchers found that exact sentence reconstruction from model updates ranged from 31% to 44%, with RadBERT demonstrating the highest fidelity in recovering clinical terms. The findings highlight that significant portions of sensitive text can be reconstructed from FL gradients, emphasizing the critical role of tokenizer design in privacy considerations for clinical NLP applications.
Tokenizer design can significantly influence the extent of privacy leakage in federated learning, with RadBERT recovering the most clinical terms despite all tokenizers failing to prevent information leakage.
Federated learning (FL) enables multi-institutional training on clinical text without sharing raw data, but gradient inversion can reconstruct sensitive information from shared model updates. The extent of this leakage for radiology reports, and the role of tokenizer design, remains unclear. We quantify gradient-based text reconstruction in FL and compare privacy risk across three tokenizers with the model architecture held fixed. Six FL clients trained a GPT-2-style transformer (sequence length 32) on public radiology corpora (368,751 diagnostic reports, 98,206 discharge summaries, 1,500 MIMIC-CXR free-text reports) using the GPT-2, RadBERT, and LLaMA-2 tokenizers at batch sizes of 64, 128, and 256. Assuming an active malicious server that modifies the shared architecture before distribution, we applied analytic gradient inversion and measured reconstruction fidelity over five runs. Exact sentence reconstruction ranged from 31% to 44% across tokenizers (30.6-43.5% across the 27 tokenizer x dataset x batch-size cells). At batch size 64 on the Discharge dataset, accuracy was 42.1% (GPT-2), 42.3% (RadBERT), and 39.4% (LLaMA-2), decreasing to 37.3%, 37.2%, and 34.3% at batch size 256. S-BLEU declined as batch size grew (GPT-2: 0.44 to 0.33; RadBERT: 0.48 to 0.35). RadBERT yielded the highest reconstruction fidelity and recovered the most clinical terms (18.1% of a 1,440-term reference vocabulary, vs 12.5% for GPT-2 and 9.4% for LLaMA-2), yet no tokenizer prevented leakage. Substantial portions of report text are therefore recoverable from FL gradients even at larger batch sizes and with domain-specific tokenizers. Tokenizer design influences leakage severity and is a privacy-relevant decision, not only a utility one; safeguards such as secure aggregation and differential privacy are likely necessary to meet HIPAA and GDPR requirements for FL in radiology NLP.
