Apppublic

Anonymized1/Interacting-With-LLMs

sourceHugging Facecc-by-4.0updated 1d agoView on Hugging Face

0likes

App README

PRISE — Privacy Risk Inference System with Explainability

PRISE is a browser-based chat interface that places a transparency layer between the user and a conversational LLM. Alongside the model's ordinary reply, each turn is annotated to show what the system has learned about the user. Identifying details in the message are highlighted and labelled by type. Phrases that match an external record about the user are marked and traced back to their source. A profile card lists the demographic attributes the model can now predict, each with its confidence and the evidence behind it. A single risk score summarises the turn.

Differential privacy can be applied on the user's side: before a message is sent, its content words are perturbed so that the model receives a noised version of the text. The conversation panel still shows what the user actually wrote, and a badge on the message reports how many words were substituted and at which privacy budget (ε).

Usage

Dependencies

Python 3.10. Install with:

bash

pip install -r requirements.txt

The main dependencies are Gradio, OpenAI, LangChain, FAISS, sentence-transformers and rank-bm25.

API keys

Keys are read from the environment; only the provider you intend to use is required.

bash

export GEMINI_API_KEY="..."         # default provider
export OPENAI_API_KEY="..."         # optional
export TOGETHER_API_KEY="..."       # optional
export VALID_ACCESS_TOKEN_1="..."   # optional, gates the app behind a token

Running

bash

python app.py

The app serves on http://localhost:7860.

Options

Every option below is available both as a command-line flag and as a URL query parameter, so a configuration can be assigned via a link:

http://localhost:7860/?rag=1&epsilon=20&show_pii_highlights=1

Option	Values	Default	Meaning
`model`	model name	`Gemini-2.5-Flash`	LLM used for the conversation
`rag`	`0` / `1`	`1`	Link the conversation against the external corpus
`epsilon`	float or `inf`	`inf`	Privacy budget for the perturbation; `inf` disables it
`show_risk`	`0` / `1`	`1`	Show the privacy risk panel
`show_tips`	`0` / `1`	`1`	Show hover tooltips
`show_pii_highlights`	`0` / `1`	`1`	Highlight identifying details
`show_rag_highlights`	`0` / `1`	`1`	Highlight matches against external data
`show_settings`	`0` / `1`	`0`	Show the live settings panel
`show_dp`	`0`–`3`	`1`	When to show the privacy settings panel: never, after each turn, at the start only, or when the conversation ends
`show_infr_attr_card`	`0`–`2`	`1`	When to show the inferred-attributes card: never, after each turn, or when the conversation ends
`enable_social_scraping`	`0` / `1`	`0`	Retrieve public social media posts as external data (experimental)
`show_social_scraping`	`0` / `1`	`0`	Show the social media control in the interface
`show_upload_data`	`0` / `1`	`0`	Show the control for uploading a custom corpus
`scenario_mode`	`real`, `persona1`, `persona2`	`real`	Free-style conversation, or a scripted persona
`demo`	`0` / `1`	`0`	Run the interface without making API calls

token is accepted as a URL parameter only, and carries the session access token. --retriever_path (an alternative retriever pickle) and --port (default 7860) are command-line flags only.

License

CC BY 4.0.