mithril-security/NonSuspiciousImageDecoder

1import gradio as gr2import cv23import io4import pandas as pd5 6from LSBSteg import LSBSteg7 8 9def convert(file):10    print(f"Converting file {file}")11    in_img = cv2.imread(file, cv2.IMREAD_UNCHANGED)12    lsbsteg = LSBSteg(in_img)13    data = lsbsteg.decode_binary()14    bytes = io.BytesIO(data)15    dataframe = pd.read_parquet(bytes)16 17    # dataframe.to_csv('output.csv')18    return dataframe.head(20)19 20 21with gr.Blocks() as demo:22    gr.Markdown("""23        ## Non-Suspicious image decoder24 25        This tool shows the extraction a dataframe hidden inside an image.26 27        There are a few ways to hide data in a PNG file, notably:28         * adding it after the end of the file (after the PNG IEND chunk), so that it gets29           ignored by image viewers30         * adding it as comments in the PNG file (tEXt chunks)31 32        These methods are kind of easy to spot! Also, a lot of software, browsers, image upload33         websites etc often just strip them.34 35        So, here, we have a different, more thoughtful (and arguably cooler) method.36 37        This class hides the data using a basic kind of **[steganography](https://en.wikipedia.org/wiki/Steganography)**:38         it hides it in the39         *least significant bits* of the raw (uncompressed) picture: tiny differences in the red, green and blue40         channel of the image encodes the data we're interested in.41 42        This means the resulting picture43         looks **very close to the original image**; and for the data we hide here, it is **imperceptible44         to the naked eye**.45 46        The resulting PNG file will probably get a little bit bigger as a result, since PNG uses compression,47         which will have a harder time when we have our stolen data injected in the image. This is48         not that much of a problem since it stays <100Ko, so it's not that noticeable.49 50        """)51    with gr.Row():52        im = gr.Image(label="Input image file", type="filepath")53 54        def preprocess(encoding: str) -> str:55            # We do our own preprocessing because gradio's deletes PNG metadata :(56            import tempfile57            import base6458 59            content = encoding.split(";")[1]60            image_encoded = content.split(",")[1]61            png_content = base64.b64decode(image_encoded)62            file_obj = tempfile.NamedTemporaryFile(63                delete=False,64                suffix=".input.png",65            )66            file_obj.write(png_content)67            return file_obj.name68 69        im.preprocess = preprocess70        df_out = gr.Dataframe(71            label="Output dataframe (first 20 rows)", max_rows=20, overflow_row_behaviour="paginate"72        )73    gr.Markdown("Click on the example below to get the data from the associated colab notebook :)")74    gr.Examples(75        examples=["sample-picture.png"],76        inputs=[im],77        outputs=[df_out],78        fn=convert,79        cache_examples=True,80    )81        # file_out = gr.File(label="Full output CSV file")82    btn = gr.Button(value="Extract")83    # demo = gr.Interface(convert, im, im_2)84    btn.click(convert, inputs=[im], outputs=[df_out])85 86    # example_img = os.path.join(os.path.dirname(__file__), "example-picture.png")87 88if __name__ == "__main__":89    demo.launch()90