The abundance of user-generated conversational data makes the question of how to responsibly derive insights from confidential data without putting the individuals in it at risk ever more important. However, the majority of existing differentially private (DP) synthetic text generation (SDG) and anonymization approaches ignore the context of the data's downstream usage, which results in loss of important information. DP-SDG provides the same guarantee regardless of the recipient, and anonymization relies on predefined attribute categories. They both ignore the observation that an attribute that is harmless to one audience can be sensitive to another, and each audience typically needs only some attributes preserved. We propose Ext-Classify-Priv, a framework that (1) extracts the information types present in a corpus, (2) classifies which of them are private for a given audience following contextual integrity, and (3) protects only those attributes, using either anonymization or SDG with Feature-DP approaches. On Reddit data with 3 audiences, our audience-specific releases are more faithful than a single audience-agnostic release for every audience. On bioRxiv data, our SDG method improves MAUVE from 0.097 to 0.420 at epsilon=1 and on the proprietary real estate data, our anonymization method improves MAUVE from 0.698 to 0.959 over the best baselines.
Figures & tables
Figure 1: Overview of Ext-Classify-Priv (private data in red, public and released data in green). (1) Extract a schema with attributes A={a1,…,ak} from the private records D . (2) Each audience Rj comes with its own release context cj , which Classify uses to split A into Apub(cj) and Apriv(cj) . The same corpus yields different splits: the advertiser R1 gets three attributes protected, the public re-post R2 only one. (3) Privatize D with respect to Apriv(cj) using SDG-CI or ANON-CI, and release D^j to Rj .
Approach
Aud.
Unit
Guarantee
Synthetic data generation
DP fine-tuning Yu et al. (2024) ; Hu et al. (2025)
✗
Record
DP
Inference DP Xie et al. (2024) ; Chien et al. (2026)
✗
Record
DP
Secret-aware DP Wang et al. (2026)
✗
Secret
DP
SDG-CI (ours)
✓
Attr.
Feat.-DP
Anonymization
Table 1: Text privatization approaches. In both categories, prior work either ignores the audience ( ✗ ) or assumes a single fixed one ( ∼ ); we add audience conditioning to each. Aud. : the protected attributes depend on who receives the data. Unit / Guarantee : what is protected and how.
Figure 2: Results of SDG ( ε=1 ) and anonymization methods. (a)–(c) Reddit released to three audiences with 6, 8, and 2 public attributes (Tab. 6 ), (d) bioRxiv with its default split (Tab. 5 ). Baselines release the same data for every audience, so only our methods change across (a)–(c). Axes, clockwise from top: MAUVE, 1/(1+KL) , 1−TV , 1−JSD , TSTR accuracy (Acc), fairness 1−∣Δ∣ (Fair, Reddit only), and privacy 1− AIA leakage (Priv). All axes are higher-is-better.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Schema construction. Top: ACTG-style prompted schema design [ 14 ] , where the schema is written from human knowledge (dataset and workload descriptions) without private data; each private document is then annotated against the schema by a trusted oracle LLM. Middle: non-private LLM baseline. Bottom: four-pass DP schema construction; only Pass 3 consumes privacy budget, and the positive threshold τ guards against releasing strings that exist only because of a single transcript, since the candidate set S(D) is itself data-dependent.
Method
ε
Fields
Signal ↑
Balanced Entropy ↑
Dominance ↓
Coverage ↑
Hand-written
n/a
47
0.144
0.33
0.72
2.19
LLM
n/a
20
0.440
0.96
0.58
3.14
MAPLE
n/a
8
0.718
0.65
0.59
2.84
DP
∞
20
0.278
0.69
0.57
3.24
DP
100
19
0.243
0.75
0.56
3.00
DP
10
15
0.236
0.63
0.65
2.65
Appendix
Table 2: Quality of the derived schemas on the RealEstate subset.
Parameter
Value
Budget
δ
1/n ( n = number of private training records)
split
ρAIM=0.1ρ , ρPE=0.9ρ
Conditional AIM
workload W
all cliques with ≤3 attributes and ≥1 private attribute
Sender: Reddit users commenting on public forums Recipient: an external ML vendor Purpose: developing a content recommendation system on users’ writing patterns
Sender: the research lab that conducted the experiments Recipient: the general public (open release) Purpose: disseminating experimental records so the findings can be read, reused and built upon
Recipient: third parties re-sharing the comments Purpose: releasing the attributes users routinely self-disclose in public posts, withholding only demographic identifiers
Imperial College London, Imperial Global Singapore · University of Manchester, United Kingdom · CFAR and IHPC, Agency for Science, Technology and Research (A*STAR), Singapore +2