1. Introduction
Owing to the increase in motorization and population, the number of traffic accidents and their victims seems to be increasing globally [
1]. Complicated traffic situations and random events pose a hazard to the safety of drivers, passengers, and pedestrians. Increasing populations and numbers of cars have made traffic accidents a major problem for transportation security. Insurance, medical, and monetary costs all go up when accidents occur on the road. Diverse factors included in traffic accidents have a significant impact on each other, consequently making it tough to individually take any of the parameters while describing the severity of traffic accidents. In the field of traffic safety research, the growth of reliable methods for predicting and classifying crash injury severity, which relies upon numerous explanatory variables, was a key factor [
2]. A mechanism for accident management serves a significant role in emergency systems and traffic control. In such structures, data from diverse sources is gathered for supporting injured people [
3].
The photographs and measurements taken at the scene of the accident are the most crucial pieces of evidence in cases of accidents. The data collected at the scene of an accident is corrected by police or investigators. There should be no room for error in accident investigations if police and investigators know exactly what they will be using the photos they take at the scene for. It is more efficient to plan out a series of high-quality images rather than taking a dozen random shots. Accident analysis relies heavily on having access to high-quality images of the incidents.
One crucial data source in accidents is image source. Portable or fixed cameras may capture such images, but the latter is very effective. Such digital images generally have a wealth of personally delicate data. When the data is analyzed and collected by attackers, unmeasurable losses will happen along with the leak of personal privacy [
4]. The privacy protection of images frequently depends on methods such as privacy encryption, k-anonymity, and access control. Several perceptual encrypted techniques were modelled to generate images without visual data according to the visual data-protection system, but data theory-related encryption (AES and RSA) creates ciphertext [
5]. Perceptual encryption intended at generating images without visual data on plain images based on a visual data-protection system as visual data involves private data such as personally identifiable information, time, and place [
6].
Conversely, there are several authors on analyzing accidents. Various image-processing approaches were advanced to get a real-time mechanism to assist the accident [
7]. Crash severity methods may forecast severity that may be anticipated to occur for a crash that aids clinics in offering proper health care as soon as possible [
8]. Moreover, research on crash injury severity even aids superior understanding of what factors contributed to injury severity once a crash occurred, which will help improve road safety and reduce crash severity. Crash severity was generally measured by numerous discrete classes of possible injury, fatal, incapacitating injury, property damage only, and non-incapacitating injury [
9].
Because of improvements in processing power and technologies, deep-learning (DL) models have achieved excellent performance in a number of domains, including autonomous vehicle systems. Now that neural networks (NN) have matured into a potent tool for discovering intricate patterns in high-dimensional datasets and delivering on-target predictions, they may now be relied upon to make accurate and trustworthy forecasts in ordinal data. Some of these techniques implement ML techniques such as artificial neural network (ANN). With the help of pooling layers, the hidden features can be derived [
10]. Generally, the output of the final pooling layer was implemented for the purposes of regression and classification.
Accident site photos and measurements are the most important evidence. Attackers will steal data and breach personal privacy, causing untold costs. The massive number of images commonly employed poses a significant challenge to privacy preservation, and image encryption can be used to accomplish cloud storage [
11] and secure image transmission in the network; moreover, an automated deep-learning (DL)-based accident severity classification is needed.
The novelty of this paper includes:
This article presents a novel Privacy Preserving Image Encryption with Optimal Deep-Learning-based Accident Severity Classification (PPIE-ODLASC) model. The goal of the presented PPIE-ODLASC technique is to accomplish secure image transmission via encryption and accident severity classification (i.e., high, medium, low, and normal).
For accident image encryption, multi-key homomorphic encryption (MKHE) technique with lion swarm optimization (LSO)-based optimal key generation process is involved.
In addition, the PPIE-ODLASC algorithm involves YOLO-v5 object detector to identify the region of interest (ROI) in the accident images. Moreover, the accident severity classification module encompasses Xception feature extractor, bidirectional gated recurrent unit (BiGRU) classification, and Bayesian optimization (BO)-based hyperparameter tuning.
The experimental validation of PPIE-ODLASC technique is tested using accident images and the results are investigated in terms of several measures.
The rest of the paper is organized as follows.
Section 2 provides a detailed review of existing models and
Section 3 elaborates the proposed algorithm. Then,
Section 4 shows experimental validation and
Section 5 draws the concluding remarks of the study.
2. Literature Review
Boulila et al. [
12] advises a hybrid PPDL method for object classification. This study aims to improve the encryption of satellite images while guaranteeing a higher object classifier accuracy and good runtime. The technique projected to encrypt the image is preserved by the public keys of somewhat homomorphic encryption and Paillier homomorphic encryption. Chuman and Kiya [
13] developed a learnable image encryption technique for privacy-preserving DNN. The presented technique is performed based on block scrambling utilized along with data augmentation methods, namely grid mask, random cropping, and horizontal flip. The usage of block scrambling improves the robustness against many attacks; on the other hand, combined with data augmentation, it allows the preservation of a higher classifier accuracy while using encrypted images.
He et al. [
14] developed a CryptoEyes to overcome the problems of privacy-preserving classifier on encrypted images. The study presents a 2-stream convolution network structure for the classifier of encrypted images to capture the contour of the encrypted image, thereby considerably increasing the accuracy of the classification. Shen et al. [
15] developed a secure SVM that is a privacy-preserving SVM training system over blockchain (BC)-based encrypted IoT information. The author utilizes the BC technique to construct reliable and secured data sharing platforms amongst various data providers, whereas an IoT information is encrypted and recorded on the distributed ledger. Ito et al. [
16] designed a transformation system to generate visually protected images for privacy-preserving DNN. However, the presented technique allows us to preserve the image classification performance and strongly protects visual information.
The authors in [
17] resolve the challenges by designing Secure DL, a privacy-preserving image detection technique for encrypted dataset over cloud. The presented block-based image encryption system is well-developed for protecting the image’s visual data. The presented technique is demonstrated to be secure from a probabilistic perspective, and with different cryptographic attacks. Ahmad and Shin [
18] present an effective pixel-based encryption technique. The technique gives a basic level of privacy while maintaining the inherent property of the original images, thus allowing DL application in the encryption field. The author has utilized logistic maps for the lower computation requirement. Furthermore, in order to compensate for any ineffectiveness due to the logistic maps, the author uses a second key for shuffling the sequence.
Li et al. [
19] proposed a new FL into autonomous driving for preserving privacy of the vehicle by sharing the model training parameter through MEC server and keeping original information in a local vehicle. Salem et al. [
20] introduce DeepZeroID: a multiple-party biometric verification and privacy-preserving cloud-based technique which makes use of homomorphic encryption. Training on sensitive biometric data is eliminated with the help of transfer learning, and one pre-trained DNN is exploited as the feature extractor. By proposing an exhaustive search algorithm, these feature extractors are employed on the processes of liveness detection and biometric authentication. Song et al. [
21] present a novel technique that constructs an effective module without sharing sensitive information between the source and target domain. The target domain benefit from the label-rich source domain without exposing its private information. Zhao et al. [
22] developed a BC based privacy-preserving software updating protocol that delivers reliable and secure updates with an incentive model while protecting the privacy of the user. Ibarrondo and Önen [
23] analyze the Batch Normalization (BN) layer: a modern layer that addresses internal covariance shift, which was demonstrated to be highly effective in improving the performance of the deep neural network. The study aims at reformulating BN that leads to a modest reduction on the number of operations in order to be compatible with the usage of FHE.
Despite the ML and DL algorithms existing in the early research, it is still necessary to optimize the privacy and accident severity classification performance. Simultaneously, various hyperparameters have a crucial effect on the effectiveness of the CNN algorithm. In particular, the hyperparameters including learning rate selection, epoch count, and batch size are necessary to attain superior outcomes. Meanwhile, the trial-and-error algorithm for hyperparameter tuning is an erroneous and challenging task; in the proposed method, the BOA algorithm was used for the parameter selection of the BiLSTM module.
4. Experimental Validation
The proposed technique is simulated by means of the Python 3.6.5 tool. The proposed model is experimented on GeForce 1050Ti 4 GB, PC i5-8600k, 16 GB RAM, 1 TB HDD, and 250 GB SSD. The parameter settings are as follows: dropout: 0.5, learning rate: 0.01, activation: ReLU, batch size: 5, and epoch count: 50. The encryption performance of the proposed model is investigated using different measures such as mean square error (MSE), PSNR, structural similarity (SSIM), and root mean square error (RMSE). Next, accuracy, precision, recall, F-score, and Mathew Correlation Coefficient (MCC) can examine the classification performance.
In this study, we examined the performance of the PPIE-ODLASC model using a set of accident images with four classes. For training purposes, we used the CADP dataset [
32], which contains 1416 video segments composed from YouTube, with 205 video segments having full spatio-temporal annotations. For testing purposes, we used our own dataset collected from a real-time environment. It comprises 20,000 samples with four classes (normal, low, medium, and high) as represented in
Table 1.
Figure 2 defines the sample images of multiclass.
Figure 3 shows the RoI extracted by the PPIE-ODLASC approach on the applied sample images. The result indicates that the PPIE-ODLASC technique has effectually extracted the RoI on all images.
Table 2 and
Figure 4 report the outcomes of the PPIE-ODLASC approach on image encryption process. The outcome stated that the PPIE-ODLASC approach has encrypted the images proficiently. For instance, on image1, the PPIE-ODLASC system has obtained an MSE of 0.1110, RMSE of 0.3332, PSNR of 57.68 dB, and SSIM of 99.81%. Meanwhile, in image3, the PPIE-ODLASC method has reached an MSE of 0.1540, RMSE of 0.3924, PSNR of 56.26 dB, and SSIM of 99.95%. Eventually, on image6, the PPIE-ODLASC technique gained an MSE of 0.1610, RMSE of 0.4012, PSNR of 56.06 dB, and SSIM of 99.87%.
Table 3 and
Figure 5 represent the PSNR results of the PPIE-ODLASC system with and without attacks. The outcome indicated that the PPIE-ODLASC algorithm has obtained effectual PSNR values under the presence of attack. For sample, in image1, the PPIE-ODLASC approach has obtained a PSNR of 57.68 dB and 56.73 dB for without and with attacks, respectively. Concurrently, on image3, the PPIE-ODLASC method has gained a PSNR of 56.26 dB and 55.14 dB for without and with attacks, correspondingly. Furthermore, in image6, the PPIE-ODLASC model has obtained a PSNR of 56.06 dB and 54.98 dB for without and with attacks, correspondingly.
A comparative PSNR study of the PPIE-ODLASC approach with other existing methods on various images is given in
Table 4 and
Figure 6. The outcome highlighted that the PPIE-ODLASC system reached higher PSNR values. For instance, in image1, the PPIE-ODLASC methodology obtained an improved PSNR of 57.68 dB, while the MSC-OKG, HSP-ECC, OGWO-ECC, and DM-CM models obtained a reduced PSNR of 55.14 dB, 51.60 dB, 48.45 dB, and 45.37 dB, respectively. Similarly, in image 3, the PPIE-ODLASC model reached an improved PSNR of 56.26 dB, while the MSC-OKG, HSP-ECC, OGWO-ECC, and DM-CM [
33] models obtained a reduced PSNR of 54.02 dB, 51.77 dB, 48.26 dB, and 45.88 dB, correspondingly. Additionally, in image 6, the PPIE-ODLASC model obtained an improved PSNR of 56.06 dB, while the MSC-OKG, HSP-ECC, OGWO-ECC, and DM-CM models obtained a reduced PSNR of 53.86 dB, 50.36 dB, 47.72 dB, and 44.69 dB, correspondingly.
The accident severity classification results of the PPIE-ODLASC model in terms of the confusion matrix are shown in
Figure 7. The results indicated that the PPIE-ODLASC model has accurately classified different types of severity levels.
Table 5 represents an overall accident severity classification result of the PPIE-ODLASC model under different sizes of TR and TS databases. The experimental results stated that the PPIE-ODLASC model has accurately identified varying levels of severity. For example, with 80% of TR data, the PPIE-ODLASC technique offered an average
of 98.32%,
of 96.68%,
of 96.65%,
of 96.65%, and MCC of 95.54%. Along with that, with 20% of TS database, the PPIE-ODLASC technique offered an average
of 98.31%,
of 96.63%,
of 96.64%,
of 96.62%, and MCC of 95.51%. Moreover, with 70% of TR database, the PPIE-ODLASC methodology offered an average
of 97.81%,
of 95.61%,
of 95.61%,
of 95.61%, and MCC of 94.15.
The TACC and VACC of the PPIE-ODLASC approach are examined on accident severity classification performance in
Figure 8. The figure exhibited that the PPIE-ODLASC method has shown improved outcomes with increased values of TACC and VACC. In particular, the PPIE-ODLASC method has reached maximum TACC outcomes.
The TLS and VLS of the PPIE-ODLASC method are tested on accident severity classification performance in
Figure 9. The figure shows that the PPIE-ODLASC approach has revealed better performance with minimal values of TLS and VLS. Notably, the PPIE-ODLASC methodology has resulted in reduced VLS outcomes.
A clear precision-recall investigation of the PPIE-ODLASC approach under test database is seen in
Figure 10. The figure indicated that the PPIE-ODLASC method has superior values of precision-recall values under several classes.
A brief ROC study of the PPIE-ODLASC method under test database is shown in
Figure 11. The result denotes the PPIE-ODLASC algorithm has demonstrated its ability in categorizing distinct classes.
In
Table 6, a detailed comparison study of the PPIE-ODLASC with current DL techniques such as CNN with multilayer perceptron (MLP), CNN with multi-kernel extreme learning machine (MELM), CNN with extreme learning machine (CNN-ELM), CNN with optimal stacked extreme learning machine (CNN-OSELM), CNN with kernel extreme learning machine (CNN-KELM), CNN with radial basis function (CNN-RBF), and CNN with SVM (CNN-SVM) is provided [
34].
Figure 12 represents the comparative accident severity classification results of the PPIE-ODLASC model with respect to
and
. The experimental results stated that the PPIE-ODLASC model has gained enhanced performance. Based on
, the PPIE-ODLASC model has gained increased
values of 96.68%, while the CNN-MLP, CNN-MELM, CNN-ELM, CNN-OSELM, CNN-KELM, CNN-RBF, and CNN-SVM models have reported reduced
values of 94.28%, 92.73%, 92.33%, 92.16%, 92.05%, 89.40%, and 88.66%, respectively. At the same time, based on
, the PPIE-ODLASC method has obtained increased
values of 96.65%, while the CNN-MLP, CNN-MELM, CNN-ELM, CNN-OSELM, CNN-KELM, CNN-RBF, and CNN-SVM [
31] approaches have reported reduced
values of 94.94%, 92.60%, 92.22%, 92.22%, 91.84%, 89.70%, and 89%, respectively.
Figure 13 represents the comparative accident severity classification results of the PPIE-ODLASC technique in terms of
and
. The result shows that the PPIE-ODLASC technique has reached enhanced performance. Based on
, the PPIE-ODLASC technique has acquired increased
values of 98.32%, while the CNN-MLP, CNN-MELM, CNN-ELM, CNN-OSELM, CNN-KELM, CNN-RBF, and CNN-SVM methods have reported reduced
values of 94.80%, 92.66%, 92.03%, 91.28%, 92.29%, 89.30%, and 86.83%, respectively.
Simultaneously, based on , the PPIE-ODLASC technique has gained increased values of 96.65%, while the CNN-MLP, CNN-MELM, CNN-ELM, CNN-OSELM, CNN-KELM, CNN-RBF, and CNN-SVM models have reported reduced values of 94.60%, 92.60%, 92.20%, 92.13%, 91.84%, 90.10%, and 88.66%, respectively.
Finally, a detailed training time (TRT) inspection of the PPIE-ODLASC with other DL methods takes place in
Figure 14. The results implied that the PPIE-ODLASC approach has gained better performance with a minimal TRT of 4.39 s. Contrastingly, the CNN-MLP, CNN-MELM, CNN-ELM, CNN-OSELM, CNN-KELM, CNN-RBF, and CNN-SVM models have reported increased TRT of 94.28%, 92.73%, 92.33%, 92.16%, 92.05%, 89.40%, and 88.66%, respectively. The result shows the superior performance of the PPIE-ODLASC approach over other existing techniques.