STAMP-R: Stylometric Text Anonymization with Memory-guided Policy Rewriting

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

STAMP-R is a risk-adaptive reinforcement learning framework designed for instance-level authorship anonymization of textual data. It addresses limitations of conventional methods that fail against stylometric inference attacks and rigid LLM-based approaches that overlook outliers, incur high costs, and offer poor privacy-utility trade-offs. Central to STAMP-R is the Style Manifold Memory (SMM), which models the global stylistic landscape using prototype-based density estimation. SMM detects high-risk stylometric outliers and adaptively modulates a composite reward function, enabling stronger obfuscation for identifiable samples while preserving semantic fidelity for low-risk instances. The framework also distills a lightweight 3B-parameter model from a teacher LLM for efficient local deployment, demonstrating reduced authorship re-identification risk and strong downstream utility.

Key takeaway

For machine learning engineers developing systems that handle sensitive textual data, STAMP-R offers a robust approach to balance privacy and utility. You should consider implementing risk-adaptive anonymization strategies that identify and target stylometric outliers, rather than using one-size-fits-all methods. This framework demonstrates how distilling lightweight models, like a 3B-parameter version, can enable efficient local deployment while effectively reducing authorship re-identification risks.

Key insights

STAMP-R provides risk-adaptive, instance-level stylometric anonymization by shaping style distributions with memory-guided policy rewriting.

Principles

• Stylometric outliers necessitate stronger, adaptive obfuscation.
• Anonymization can be formulated as instance-level style distribution shaping.
• Adaptive obfuscation improves the privacy-utility trade-off.

Method

STAMP-R formulates anonymization as a risk-aware, instance-level style distribution shaping problem. It uses Style Manifold Memory (SMM) for prototype-based density estimation to detect high-risk stylometric outliers, adaptively modulating a composite reward function. A lightweight 3B-parameter model is distilled for efficient deployment.

In practice

• Implement prototype-based density estimation to identify stylometric outliers.
• Apply adaptive obfuscation strategies based on instance-level risk.
• Distill large models into lightweight versions (e.g., 3B parameters) for efficient deployment.

Topics

• Stylometric Anonymization
• Reinforcement Learning
• Large Language Models
• Privacy-Utility Trade-off
• Text Obfuscation
• Style Manifold Memory

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Privacy-Preserving Text Sanitization for Distributed Agents Collaboration via Disentangled Representations — Privacy leakage in multi-agent text collaboration extends beyond explicit identifiers to stylistic patterns, necessitating disentangled sanitization.
• MultiGraSCCo: A Multilingual Anonymization Benchmark with Annotations of Personal Identifiers — Synthetic, culturally-adapted multilingual datasets can overcome patient data scarcity for privacy-preserving AI development.
• How probability models protect privacy — Differential privacy uses calibrated randomness to protect individual data while enabling accurate aggregate analysis.
• Seeing Without Exposing: Adaptive Privacy Control for Open-World, Context-Hungry MLLMs — Adaptive privacy drifting for MLLMs protects sensitive data by semantically altering elements while preserving essential visual context.
• Hybrid privacy-aware semantic search: SVD-truncated document geometry and CKKS-encrypted query reranking under a restricted threat model — Hybrid semantic search protects documents via SVD truncation and queries via CKKS encryption, balancing privacy and performance.
• Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models — Distribution shifts critically impact practical privacy in differentially private LLM adaptations, often undermining theoretical guarantees.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.