PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepKeyGen: A Deep Learning-based Stream Cipher Generator for Medical Image Encryption and Decryption

Yi Ding, Fuyuan Tan, Zhen Qin, Mingsheng Cao, Kim-Kwang Raymond Choo, Zhiguang Qin

arXiv:2012.11097v1cs.CRcs.AIcs.NE

TL;DR

Medical-image leakage can expose sensitive patient information, motivating secure key generation. DeepKeyGen uses a GAN to translate a seed image into a private key guided by a transformation domain. Evaluations on three medical-image datasets report high security for the generated stream cipher and good performance in encrypting multimodality images.

  • Problem

    Medical images contain sensitive patient information, and leakage can have privacy implications and legal ramifications for hospitals.

    Source: §I · p. 1

  • Method

    DeepKeyGen uses a GAN to translate an initial source-domain image into a private key, with the transformation domain specifying the desired key style.

    Source: §I · p. 2, §VI · p. 14

  • Results

    Across three medical-image datasets, the authors report that DeepKeyGen's stream cipher has a large key space, pseudo-randomness, one-time-pad properties, sensitivity to change, and resistance to different attacks.

    Source: §VI · p. 15

  • Takeaways & Limitations

    The authors report that DeepKeyGen achieves a high level of security compared with other key-generation algorithms and good performance encrypting multimodality medical images.

    Source: §VI · p. 15

    Future work will investigate lightweight networks to improve efficiency and evaluate security and performance in other application domains to assess generalizability.

    Source: §VI · p. 15

Abstract

from arXiv · show

The need for medical image encryption is increasingly pronounced, for example to safeguard the privacy of the patients' medical imaging data. In this paper, a novel deep learning-based key generation network (DeepKeyGen) is proposed as a stream cipher generator to generate the private key, which can then be used for encrypting and decrypting of medical images. In DeepKeyGen, the generative adversarial network (GAN) is adopted as the learning network to generate the private key. Furthermore, the transformation domain (that represents the "style" of the private key to be generated) is designed to guide the learning network to realize the private key generation process. The goal of DeepKeyGen is to learn the mapping relationship of how to transfer the initial image to the private key. We evaluate DeepKeyGen using three datasets, namely: the Montgomery County chest X-ray dataset, the Ultrasonic Brachial Plexus dataset, and the BraTS18 dataset. The evaluation findings and security analysis show that the proposed key generation network can achieve a high-level security in generating the private key.

I. INTRODUCTION

Medical-image privacy motivates secure key generation, but existing generators require costly manual design and refinement. DeepKeyGen instead learns to generate private keys through image-to-image translation, with the transformation domain specifying the desired key style.

I. INTRODUCTION

  • Medical images contain sensitive patient information, and their leakage can have privacy implications and legal ramifications for hospitals.The introduction motivates security solutions, including cryptographic primitives, to protect these images and patient privacy.Source: §I · p. 1
  • DeepKeyGen is proposed to learn private-key generation through image-to-image translation rather than rely on manually designed generators.The paper describes existing generator design and refinement as time- and resource-expensive.Source: §I · pp. 1–2, A. Existing Key Generation Algorithms · p. 3
  • The transformation domain represents the desired private-key style, while a source-domain image serves as the seed that the GAN translates into a key.The discriminator distinguishes generated keys from data in the transformation domain; training uses unlabeled, unpaired images.Source: §I · p. 2
  • In the example healthcare workflow, a generated key encrypts a medical image, and the ciphertext and key are stored in PACS for later retrieval and decryption.The described system assumes an intranet environment and secure transmission channels.Source: §I · p. 2
  • The authors report evaluating DeepKeyGen on chest X-ray, ultrasonic brachial plexus, and BraTS18 datasets, with findings of high security and randomness, efficient encryption, and resilience to attacks.The introduction also states that the evaluations demonstrate generality across the three datasets.Source: §I · p. 2

A. Existing Key Generation Algorithms

  • Existing key-generation approaches include ECG-feature, power-system-state, DNA/deep-learning, and chaotic-system methods, but their generators require repeated manual design and refinement.The authors characterize this process as time- and resource-expensive.Source: A. Existing Key Generation Algorithms · pp. 2–3

B. Deep Learning-based Image-to-Image Transformation

Image-to-image translation learns mappings between image domains, and GANs use a generator and discriminator in that process. DeepKeyGen applies this approach to private-key generation and combines the generated stream cipher with XOR encryption and decryption.

B. Deep Learning-based Image-to-Image Transformation

  • GANs pair a sample-generating network with a discriminator, and image-to-image translation methods map images between domains; only a few researchers had applied deep learning to private-key generation.The paper identifies unpaired and unlabeled data as useful for image translation methods.Source: B. Deep Learning-based Image-to-Image… · p. 3

A. Encryption and Decryption Architecture

  • The system encrypts plaintext with the generated private key using XOR, producing ciphertext, and reverses the operation for decryption.The transformation-domain variable y is used as ground truth in the discriminator network.Source: A. Encryption and Decryption… · p. 3

B. DeepKeyGen

  • DeepKeyGen generates a private key by translating a seed image from the source domain into the transformation domain, which represents the desired key style.Its generator produces the key, while the discriminator compares generated keys with transformation-domain data; the training loss includes generator and discriminator losses.Source: B. DeepKeyGen · p. 4
  • DeepKeyGen comprises a generator that creates keys from initial images and a discriminator that distinguishes generated keys from transformation-domain data.The stated training objective combines the generator and discriminator loss functions.Source: B. DeepKeyGen · pp. 3–4

1) Generator Network G:

DeepKeyGen uses an adversarial generator to map seed images toward the transformation-domain style, producing an image-form private key. The key represents each image pixel together with its spatial and color-channel information.

1) Generator Network G:

  • The generator maps seed images toward the transformation domain, producing private keys with its style; its layers extract features, build content and diverse features, then reconstruct an image.Instance normalization is used to improve generated-image quality, accelerate convergence, and prevent gradient explosion.Source: 1) Generator Network G · p. 4

2) Discriminator Network D:

  • The discriminator classifies generated images as real or fake relative to the transformation domain, while its adversarial loss maximizes classification accuracy.At balance, its accuracy should be close to 50%, meaning it cannot distinguish generated keys from transformation-domain data.Source: 2) Discriminator Network D · p. 4

C. Image Private Key

  • The private key is an image represented as a sequence of pixels, with each pixel carrying its value, horizontal and vertical coordinates, and RGB channel.The authors describe this four-dimensional key as providing a sufficiently large key space and enhancing security.Source: C. Image Private Key · p. 4

D. Key Generation Process

Key generation trains DeepKeyGen to learn a mapping from the source domain to the transformation domain. Forward generation and discriminator evaluation contribute losses that are backpropagated to update network parameters.

D. Key Generation Process

  • DeepKeyGen learns the source-to-transformation-domain mapping by training a generator and discriminator, with randomly initialized convolutional parameters.The generator takes an initial image as input and produces the private key.Source: D. Key Generation Process · pp. 4–5
  • The process initializes the network, trains the generator and discriminator simultaneously, then updates both networks’ parameters until generation performance is best.The algorithm converts source and transformation-domain images into 256 × 256 × 3 matrices and combines the two losses.Source: Figure 3 · p. 5, D. Key Generation Process · p. 5
  • Forward propagation generates a key from the seed, and the combined generator and discriminator losses are backpropagated to update network parameters by gradient descent.The generator loss measures closeness to the target key, and the networks are trained in an alternating manner until the epoch target or stable loss is reached.Source: D. Key Generation Process · p. 5

E. Imitation Learning Attack Models

The paper considers imitation-learning attacks under different combinations of leaked model information and describes the most exposed scenario as requiring one-time-pad-like key variation. Its security evaluation uses medical-image datasets and XOR encryption.

E. Imitation Learning Attack Models

  • Imitation learning assumes an attacker can train a similar network if DeepKeyGen’s structure or transformation domain is leaked, with three scenarios spanning either or both disclosures.The paper calls networks learned by imitation “attack models.”Source: E. Imitation Learning Attack Models · pp. 5–6
  • The generation process initializes generator and discriminator networks, runs seed-based key generation alongside discriminator evaluation, then backpropagates to update both networks.Generator and discriminator training proceed simultaneously before the backward update.Source: Figure 4 · p. 6

1) Transformation Domain Leakage:

  • With only the transformation domain leaked, an attacker trains a differently structured network on that domain and tests whether its key decrypts ciphertext encrypted using DeepKeyGen’s key.Successful decryption is treated as cracking the private key through imitation learning.Source: 1) Transformation Domain Leakage · p. 6

2) Network Structure Leakage:

  • With only the network structure exposed, the attacker trains that structure using different transformation domains and tests whether its generated key decrypts the ciphertext.Successful decryption counts as cracking the key.Source: 2) Network Structure Leakage · p. 6

3) Transformation Domain and Network Structure Leakage:

IV. SECURITY ANALYSIS

  • The evaluation combines DeepKeyGen keys with XOR encryption and uses Montgomery County chest X-rays as seeds, with 138 images in each of the source and transformation domains.The evaluation also uses the Ultrasonic Brachial Plexus and BraTS18 datasets; training uses 256 × 256 inputs, batch size 1, and 20,000 epochs.Source: §IV · p. 6

A. Key Security Analysis

DeepKeyGen’s image-form private keys are described as having a large key space, varying across separate training runs, and showing high entropy. The section evaluates these properties as indicators of key security.

1) Key Space:

  • The 256 × 256 × 3 image key, with pixel values from 0 to 255, is described as having a key space sufficiently large to resist exhaustive attacks.The passage gives the key-space expression as (28)196608.Source: 1) Key Space · p. 7

2) One-Time Pad:

  • Four models trained under the same settings produce visually different keys from the same seed, supporting DeepKeyGen’s characterization as a type of one-time pad.The reported differences include key-image content, color, and contour.Source: 2) One-Time Pad · p. 7, Figure 5 · p. 7

3) Information Entropy:

4) Histogram Analysis:

The private keys have relatively uniform histograms, and changing one seed pixel produces substantial differences between generated keys. Across the three medical-image datasets, ciphertext entropy is close to 8.

4) Histogram Analysis:

  • The histogram of a generated private key is described as relatively uniform, with pixel-value frequencies close to one another.The authors interpret this distribution as high randomness that can resist statistical attacks.Source: 4) Histogram Analysis · p. 7, Figure 6 · p. 8

5) Sensitivity Analysis:

  • A one-pixel change to the initial image results in over 99.5% differences between generated keys and average intensity changes above 20%.The authors interpret this as sensitivity to the seed, contributing to pseudo-randomness and uncertainty.Source: Table IV · p. 8, 5) Sensitivity Analysis · p. 8

1) Information Entropy Analysis:

  • 7.9986, 7.9969, and 7.9939 are the average ciphertext entropies for the chest X-ray, brachial plexus, and BraTS18 datasets, respectively.The corresponding plaintext averages are 7.7360, 7.3878, and 5.8041; the paper describes 8 as the theoretical optimum and concludes that plaintext statistical information is protected.Source: Table V · p. 8, Table VI · p. 8, Table VII · p. 8

2) Histogram Analysis:

Encrypted-image histograms are described as relatively uniform and unlike those of the plaintexts across the three datasets. MSE and SSIM results also indicate low plaintext–ciphertext similarity across the datasets.

2) Histogram Analysis:

  • Encrypted-image pixel distributions are relatively uniform and close to white-noise histograms across the chest X-ray, brachial plexus, and BraTS18 datasets.The authors state that the encrypted images protect plaintext statistical information and make useful information difficult to acquire through statistical attacks.Source: 2) Histogram Analysis · pp. 8–9, Figure 7 · p. 9

3) Similarity Analysis:

  • 10577.68, 11946.85, and 9741.57 are the average MSE values, while 0.0020, 0.0047, and 0.0021 are the SSIM values for the X-ray, brachial plexus, and BraTS18 datasets, respectively.The authors interpret the large MSE values and SSIM values close to 0 as indicating low similarity between plaintext and ciphertext images.Source: Table VIII · p. 9, Table IX · p. 9, Table X · p. 9

4) Correlation Analysis:

The paper evaluates adjacent-pixel correlation in plaintext and encrypted images using samples from horizontal, vertical, and diagonal directions. It reports that encrypted-image correlations are almost zero, unlike the correlations present in plaintext images.

4) Correlation Analysis:

  • Encrypted-image adjacent-pixel correlations are almost zero, whereas plaintext images retain some correlation, across the three datasets.The analysis uses 256 adjacent-pixel pairs in horizontal, vertical, and diagonal directions; the authors state this can resist attacks exploiting strong correlation.Source: 4) Correlation Analysis · pp. 9–10, Table XI · p. 10, Table XII · p. 10

C. Traditional Attack Models

The attack evaluations report resistance to ciphertext-only, known-plaintext, chosen-plaintext, and chosen-ciphertext attacks. Separate leakage tests find that knowing the transformation domain alone does not let an attacker reproduce a working key.

1) Ciphertext Only Attack:

  • The reported key space, (28)196608, and high randomness make the generated key challenging to guess in a ciphertext-only attack.The authors also report a complex key-generation process.Source: 1) Ciphertext Only Attack · p. 10

2) Known-plaintext Attack:

  • Low plaintext-ciphertext similarity, high ciphertext entropy, and low ciphertext correlation support the reported resistance to known-plaintext attacks.The authors conclude that contextual analysis cannot recover the complete plaintext from the ciphertext.Source: 2) Known-plaintext Attack · p. 10

3) Chosen-plaintext Attack / Differential Attack:

4) Chosen-ciphertext Attack:

D. Security Against Imitation Learning Attackers

1) Transformation Domain Leakage:

  • When the attacker knows the transformation domain, keys from alternative network structures fail to decrypt the ciphertext, while key A restores the image.The figure compares plaintext, ciphertext, and decryptions using keys A–D.Source: 1) Transformation Domain Leakage · p. 11, Figure 8 · p. 11

2) Network Structure Leakage:

The leakage experiments find that different trained networks or transformation domains produce keys that do not mutually decrypt ciphertexts. Comparisons with existing key-generation algorithms report stronger randomness and a larger key space for DeepKeyGen, alongside a roughly two-second generation time.

2) Network Structure Leakage:

  • A key generated with transformation domain B cannot decrypt ciphertext encrypted with a key generated using domain A.The experiments keep network structure and other settings the same while changing the transformation domain.Source: 2) Network Structure Leakage · p. 11, Figure 9 · p. 11

3) Transformation Domain and Network Structure Leakage:

E. Comparison with Existing Key Generation Algorithms

V. PERFORMANCE EVALUATION

The hyperparameter evaluation finds the strongest reported key-entropy result with a learning rate of 0.0002, 20,000 epochs, and batch size 1. Larger learning rates or batch sizes can hinder key generation or training convergence.

A. Hyper-parameters

  • Batch size 1 prevents key generation in most epochs, while batch sizes 6 or 10 are associated with network collapse despite more training epochs.The authors interpret these outcomes as training non-convergence or collapse.Source: A. Hyper-parameters · p. 13
  • 9.9798 information entropy is the reported best result, with learning rate 0.0002, 20,000 training epochs, and batch size 1.The evaluation uses the Montgomery County chest X-ray dataset, split into 90% training and 10% validation data.Source: A. Hyper-parameters · p. 13, Table XXI · p. 13
  • At learning rate 0.002, generated-key entropy remains below the stated security expectation of about 7.9, while entropy improves as the learning rate decreases.The authors report that most entropy values exceed 7.9 at learning rate 0.0002.Source: A. Hyper-parameters · p. 13

B. Performance of Image Encryption and Decryption

Across three medical-image datasets, DeepKeyGen generates keys that encrypt images into visually different ciphertexts and restore them during decryption. The results also show that encryption performance depends on the security of the chosen transformation domain.

B. Performance of Image Encryption and Decryption

C. Impact on Transformation Domain

  • With low-security domain A, key entropy is 7.8713 and ciphertext entropy is 7.9865; with high-security domain B, the respective values are 7.9798 and 7.9986.The chest outline is exposed in ciphertext using domain A, while domain B protects the medical information better.Source: C. Impact on Transformation Domain · p. 14, Figure 13 · p. 14

VI. CONCLUSION

  • DeepKeyGen learns a mapping from a seed image to a private key by translating the seed toward the transformation domain’s desired style.The generated key is used with XOR to encrypt and decrypt medical images.Source: §VI · pp. 14–15
  • The authors identify lightweight networks as future work to improve efficiency and plan to evaluate performance and security in other application domains.They also plan to release the prototype and related materials.Source: §VI · p. 15
Loading 2012.11097v1…