|
|
import gradio as gr |
|
|
from models.watermark_faster import watermark_model |
|
|
from options import get_parser_main_model |
|
|
|
|
|
opts = get_parser_main_model().parse_args() |
|
|
model = watermark_model(language=opts.language, mode=opts.mode, tau_word=opts.tau_word, lamda=opts.lamda) |
|
|
|
|
|
def create_model(language,tau_word): |
|
|
global model |
|
|
model = watermark_model(language=language, mode=opts.mode, tau_word=tau_word, lamda=opts.lamda) |
|
|
|
|
|
return language,tau_word |
|
|
def watermark_embed_demo(raw,tau_word): |
|
|
|
|
|
watermarked_text = model.embed(raw,tau_word) |
|
|
return watermarked_text |
|
|
|
|
|
def watermark_extract(raw): |
|
|
is_watermark, p_value, n, ones, z_value = model.watermark_detector_fast(raw) |
|
|
confidence = (1 - p_value) * 100 |
|
|
|
|
|
return f"{confidence:.2f}%" |
|
|
|
|
|
def precise_watermark_detect(raw,tau_word): |
|
|
is_watermark, p_value, n, ones, z_value = model.watermark_detector_precise(raw,tau_word) |
|
|
confidence = (1 - p_value) * 100 |
|
|
|
|
|
return f"{confidence:.2f}%" |
|
|
|
|
|
demo = gr.Blocks() |
|
|
with demo: |
|
|
with gr.Column(): |
|
|
with gr.Row(): |
|
|
with gr.Column(scale=9): |
|
|
gr.Markdown( |
|
|
""" |
|
|
# 💦[Watermarking Text Generated by Black-Box Language Models](https://arxiv.org/abs/2305.08883) |
|
|
""" |
|
|
) |
|
|
language = gr.Dropdown( |
|
|
label="Language", choices=["English", "Chinese"], value="English" |
|
|
) |
|
|
tau_word = gr.Number(label="tau_word", value=0.8) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
with gr.Tab("Welcome"): |
|
|
gr.Markdown( |
|
|
""" |
|
|
This space exhibits a watermarking technique that allows third parties to independently inject an authentication watermark into generated text. |
|
|
We provide implementations for both English and Chinese text (you can select the respective language in the top right corner). |
|
|
Furthermore, you can adjust the value of $\\tau_{word}$ to control the similarity between the original text and the watermarked text. |
|
|
We recommend setting $\\tau_{word}$ at 0.8 for English and 0.75 for Chinese. |
|
|
Generally, a larger $\\tau_{word}$ increases the similarity between the original and watermarked text, but it also weakens the strength of the watermark. |
|
|
More details can be found in our [ArXiv preprint](https://arxiv.org/abs/2305.08883). |
|
|
""" |
|
|
) |
|
|
gr.Markdown( |
|
|
""" |
|
|
For a better experience, you may duplicate the space and upgrade to GPU in settings. |
|
|
""" |
|
|
) |
|
|
with gr.Tab("Watermark Injection & Detection"): |
|
|
language.change(fn=create_model, inputs=language,outputs=language) |
|
|
with gr.Row(): |
|
|
inputs = gr.TextArea(label="Input text", placeholder="Copy your text here...") |
|
|
output = gr.Textbox(label="Watermarked Text",lines=7) |
|
|
analysis_button = gr.Button("Inject Watermark") |
|
|
inputs_embed = [inputs,tau_word] |
|
|
analysis_button.click(fn=watermark_embed_demo, inputs=inputs_embed, outputs=output) |
|
|
|
|
|
inputs_w = gr.TextArea(label="Text to Analyze", placeholder="Copy your watermarked text here...") |
|
|
with gr.Row(): |
|
|
mode = gr.Dropdown( |
|
|
label="Detection Mode", choices=["Fast", "Precise"], value="Fast" |
|
|
) |
|
|
output_detect = gr.Textbox(label="Confidence (the likelihood of the text containing a watermark)") |
|
|
detect_button = gr.Button("Detect") |
|
|
|
|
|
def detect_watermark(inputs_w, mode, tau_word): |
|
|
if mode == "Fast": |
|
|
return watermark_extract(inputs_w) |
|
|
else: |
|
|
return precise_watermark_detect(inputs_w,tau_word) |
|
|
|
|
|
detect_button.click(fn=detect_watermark, inputs=[inputs_w, mode, tau_word], outputs=output_detect) |
|
|
|
|
|
if __name__ == "__main__": |
|
|
gr.close_all() |
|
|
demo.title = "Watermarking Text Generated by Black-Box Language Models" |
|
|
demo.launch() |
|
|
|
