Skip to content

PixinWavICASSP2022

stegobox.codec.PixinWavICASSP2022

Bases: BaseCodec

PixInWav targets the multimodal case of hiding images in audio.

To this end, we propose a novel residual architecture operating on top of short-time discrete cosine transform (STDCT) audio spectrograms. Among our results, we find that the residual steganography setup we propose allows an encoding of the hidden image that is independent from the host audio without compromising quality.

This repository includes a python implemenation of StegoUNet, a deep neural network modelling an audio steganographic function.

  • Created by: QiuYu
  • Created time: 2022/11/08

This is a PyTorch implementation of image steganography via deep learning, which is released in paper - Pixinwav: Residual Steganography for Hiding Pixels in Audio.

Originally implemented in margaritageleta/PixInWav.

Source code in stegobox/codec/pixinwav_icassp2022/pixinwav_icassp2022.py 
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
class PixinWavICASSP2022(BaseCodec):
    """PixInWav targets the multimodal case of hiding images in audio.

    To this end, we propose a novel residual architecture operating on top of short-time
    discrete cosine transform (STDCT) audio spectrograms. Among our results, we find
    that the residual steganography setup we propose allows an encoding of the hidden
    image that is independent from the host audio without compromising quality.

    This repository includes a python implemenation of StegoUNet, a deep neural network
    modelling an audio steganographic function.

    * Created by: QiuYu
    * Created time: 2022/11/08

    This is a PyTorch implementation of image steganography via deep learning, which is
    released in paper - [Pixinwav: Residual Steganography for Hiding Pixels in
    Audio](https://ieeexplore.ieee.org/document/9746191/).

    Originally implemented in
    [margaritageleta/PixInWav](https://github.com/margaritageleta/PixInWav).
    """

    def __init__(self, weights: str = "ckpt/pixinwav/checkpoint.pt") -> None:
        super().__init__()
        self.model = StegoUNet(
            # architecture=args.architecture,
            transform="cosine",
            add_noise=False,
            noise_kind=None,
            noise_amplitude=None,
        )

        self.checkpoint = torch.load(weights, map_location="cpu")
        self.model = nn.DataParallel(self.model)  # type: ignore
        self.model.load_state_dict(self.checkpoint["state_dict"], False)
        self.payload = ""
        print("Checkpoint loaded ++")
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.model.to(self.device)

    def imagetransform(
        self, image: Image.Image, colorspace="RGB", n=256, proportion=2**6
    ):
        image2 = image.convert(colorspace)
        nx, ny = image2.size  # type: ignore
        n = min(nx, ny)
        left = top = n / proportion
        right = bottom = (proportion - 1) * n / proportion
        image3 = image2.crop((left, top, right, bottom))  # type: ignore
        image4 = image3.resize((256, 256), Image.ANTIALIAS)  # type: ignore
        image5 = np.array(image4).astype("float") / 255.0
        return image5

    def audiotransform(self, transform, sound):
        _limit = 67522  # 2 ** 16 + 2 ** 11 - 2 ** 6 + 2
        _frame_length = 2**12 if transform == "cosine" else 2**11 - 1
        _frame_step = 2**6 - 2 if transform == "cosine" else 132

        _transform = transform
        if _transform == "fourier":
            stft = STFT(
                filter_length=_frame_length,
                hop_length=_frame_step,
                win_length=_frame_length,
                window="hann",
            )
        sound = sound  # type: ignore
        # Get the samples dimension
        sound = sound[0]
        # Create a temporary array
        tmp = torch.zeros(
            [
                _limit,
            ]
        ).normal_(mean=0, std=0.005)
        # Cut the audio on limit
        if sound.numel() < _limit:
            tmp[: sound.numel()] = sound[:]
        else:
            i = random.randint(0, len(sound) - _limit)
            tmp[:] = sound[i : i + _limit]
        if _transform == "cosine":
            return sdct_torch(
                tmp.type(torch.float32),
                frame_length=_frame_length,
                frame_step=_frame_step,
            )
        elif _transform == "fourier":
            magnitude, phase = stft.transform(  # type: ignore
                tmp.unsqueeze(0).type(torch.float32)
            )
            return magnitude, phase

        else:
            raise Exception("Transform not implemented")

    def encode(self, carrier: torch.Tensor, payload: Image.Image) -> torch.Tensor:
        """Encoder requires carrier audio to be WAV and payload to be a image.

        Args:
            carrier: Carrier audio in format WAV.
                     Read with `stegobox.io.torchaudiofile.read()`.
            payload: Payload (secret message) to be encoded. Payload is a image.
                     Read with `stegobox.io.txt.read_bytes()`

        Returns:
            container_wav: Can be formed to container audio through
                           'io.torchaudiofile.write()' function
        """
        carrier2 = self.audiotransform("cosine", carrier)
        carrier2.to(self.device)  # type: ignore
        payload2 = np.asarray(self.imagetransform(payload, n=256)).astype("float64")
        payload3 = torch.from_numpy(payload2).unsqueeze(0)
        payload4 = (
            payload3.permute(0, 3, 1, 2)
            .type(torch.cuda.FloatTensor)  # type: ignore
            .to(self.device)  # type: ignore
        )
        container, self.payload = self.model(payload4, carrier2)
        container_wav = isdct_torch(
            container.squeeze(0).squeeze(0),
            frame_length=4096,
            frame_step=62,
            window=torch.hamming_window,
        )
        return container_wav

    def decode(self, carrier: torch.Tensor) -> torch.Tensor:
        """Decode the secret payload from the carrier audio

        Args:
            carrier: Carrier audio in format WAV.
                     Read with `stegobox.io.torchaudiofile.read()`.

        Returns:
            reveal: The decoded payload image (secret message).
                Can be formed to image through 'io.image.write_tensor()' function
        """
        carrier2 = self.audiotransform("cosine", carrier)
        carrier3 = carrier2.to(self.device)  # type: ignore
        carrier3 = carrier3.unsqueeze(0)
        carrier3 = carrier3.unsqueeze(0)
        print("xxxxxxxxxx", carrier3.size())
        reveal = self.model.module.RN.forward(carrier3)  # type: ignore
        return reveal

encode(carrier, payload)

Encoder requires carrier audio to be WAV and payload to be a image.

Parameters:

Name Type Description Default
carrier Tensor

Carrier audio in format WAV. Read with stegobox.io.torchaudiofile.read().

 required
payload Image

Payload (secret message) to be encoded. Payload is a image. Read with stegobox.io.txt.read_bytes()

 required

Returns:

Name Type Description
container_wav Tensor

Can be formed to container audio through 'io.torchaudiofile.write()' function
Source code in stegobox/codec/pixinwav_icassp2022/pixinwav_icassp2022.py 
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
def encode(self, carrier: torch.Tensor, payload: Image.Image) -> torch.Tensor:
    """Encoder requires carrier audio to be WAV and payload to be a image.

    Args:
        carrier: Carrier audio in format WAV.
                 Read with `stegobox.io.torchaudiofile.read()`.
        payload: Payload (secret message) to be encoded. Payload is a image.
                 Read with `stegobox.io.txt.read_bytes()`

    Returns:
        container_wav: Can be formed to container audio through
                       'io.torchaudiofile.write()' function
    """
    carrier2 = self.audiotransform("cosine", carrier)
    carrier2.to(self.device)  # type: ignore
    payload2 = np.asarray(self.imagetransform(payload, n=256)).astype("float64")
    payload3 = torch.from_numpy(payload2).unsqueeze(0)
    payload4 = (
        payload3.permute(0, 3, 1, 2)
        .type(torch.cuda.FloatTensor)  # type: ignore
        .to(self.device)  # type: ignore
    )
    container, self.payload = self.model(payload4, carrier2)
    container_wav = isdct_torch(
        container.squeeze(0).squeeze(0),
        frame_length=4096,
        frame_step=62,
        window=torch.hamming_window,
    )
    return container_wav

decode(carrier)

Decode the secret payload from the carrier audio

Parameters:

Name Type Description Default
carrier Tensor

Carrier audio in format WAV. Read with stegobox.io.torchaudiofile.read().

 required

Returns:

Name Type Description
reveal Tensor

The decoded payload image (secret message). Can be formed to image through 'io.image.write_tensor()' function
Source code in stegobox/codec/pixinwav_icassp2022/pixinwav_icassp2022.py 
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
def decode(self, carrier: torch.Tensor) -> torch.Tensor:
    """Decode the secret payload from the carrier audio

    Args:
        carrier: Carrier audio in format WAV.
                 Read with `stegobox.io.torchaudiofile.read()`.

    Returns:
        reveal: The decoded payload image (secret message).
            Can be formed to image through 'io.image.write_tensor()' function
    """
    carrier2 = self.audiotransform("cosine", carrier)
    carrier3 = carrier2.to(self.device)  # type: ignore
    carrier3 = carrier3.unsqueeze(0)
    carrier3 = carrier3.unsqueeze(0)
    print("xxxxxxxxxx", carrier3.size())
    reveal = self.model.module.RN.forward(carrier3)  # type: ignore
    return reveal