Image classification models have demonstrated remarkable performance across various applications, yet they remain vulnerable to adversarial attacks, which can significantly impair their accuracy and reliability. This ...
详细信息
ISBN:
(数字)9798331533137
ISBN:
(纸本)9798331533144
Image classification models have demonstrated remarkable performance across various applications, yet they remain vulnerable to adversarial attacks, which can significantly impair their accuracy and reliability. This paper presents a new hybrid defense scheme to increase adversarial robustness in image classifiers. The approach combines two complementary techniques: adversarial training and input transformations. Adversarial training is implemented using the Projected Gradient Descent (PGD) attack to generate robust features by exposing the model to adversarial samples during training. Concurrently, input transformations, including random resizing, JPEG compression, and noise injection, disrupt adversarial perturbations and preserve critical image features. The integration of these methods results in a multi-layered defense mechanism that improves the model's resilience to a range of adversarial attacks. Extensive experiments are conducted on standard image classification datasets, evaluate the efficacy of the proposed scheme against various attack methods, including FGSM and PGD. The results demonstrate that the hybrid defense strategy significantly enhances robustness while maintaining competitive performance on clean images. This research offers a comprehensive solution for improving the reliability of image classifiers in adversarial settings and provides insights into balancing robustness and accuracy.
We present the ***-VBR winning candidate codec recently selected by Question 9 of Study Group 16 (Q9/16) of ITU-T as a baseline for the development of a scalable solution for wideband speech and audio compression at r...
详细信息
We present the ***-VBR winning candidate codec recently selected by Question 9 of Study Group 16 (Q9/16) of ITU-T as a baseline for the development of a scalable solution for wideband speech and audio compression at rates between 8 kb/s and 32 kb/s. The Q9/16 codec is an embedded codec comprising 5 layers where higher layer bitstreams can be discarded without affecting the decoding of the lower layers. The two lower layers are based on the CELP technology where the core layer takes advantage of signal classification based encoding. The higher layers encode the weighted error signal from lower layers using overlap-add transform coding. The codec has been designed with the primary objective of a high-performance wideband speech coding for error- prone telecommunications channels, without compromising the quality for narrowband/wideband speech or wideband music signals. The codec performance is demonstrated with selected test results.
In this paper we investigated how merging of MPEG-7 visual descriptors can improve accuracy of image classification systems. Combination of MPEG-7 Colour Layout Descriptor (CLD) and Edge Histogram Descriptor (EHD) was...
详细信息
In this paper we investigated how merging of MPEG-7 visual descriptors can improve accuracy of image classification systems. Combination of MPEG-7 Colour Layout Descriptor (CLD) and Edge Histogram Descriptor (EHD) was used to extract features for input into multiclass Support Vector Machine (SVM) classifier. SVM classifier was trained and evaluated on a standard test database of 1000 images classified into 10 categories. Achieved accuracy for merged descriptors has shown significant improvement compared to the results when CLD and EHD descriptors were used independently.
The objective of this research is to design a new JPEG-based compression scheme which simultaneously considers the security issue. Our method starts from dividing image into non-overlapping blocks with size 8×8. ...
详细信息
The objective of this research is to design a new JPEG-based compression scheme which simultaneously considers the security issue. Our method starts from dividing image into non-overlapping blocks with size 8×8. Among these blocks, some are used as reference blocks and the rest are used as query blocks. A query block is the combination of the residual and the resultant of a filtered reference block. We put our emphasis on how to estimate an appropriate filter and then use it as part of a secret key. With both reference blocks and the residuals of query blocks, one is able to encode secured images using a correct secret key. The experiment results will demonstrate that how different secret keys can control the quality of restored image based on the priority of authority.
This paper presents a low-complexity video encoding method applicable for wireless image transmission in capsule endoscopes. This encoding method is based on Wyner-Ziv theory, in which side information available at a ...
详细信息
This paper presents a low-complexity video encoding method applicable for wireless image transmission in capsule endoscopes. This encoding method is based on Wyner-Ziv theory, in which side information available at a transmitter is treated as side information at its receiver. Therefore complex processes in video encoding, such as estimation of the motion vector, are moved to the receiver side, which has a larger-capacity battery. As a result, the encoding process is only to decimate coded original data through channel coding. We provide a performance evaluation for a low-density parity check (LDPC) coding method in the AWGN channel.
The paper presents an efficient algorithm for post compression optimal rate allocation and packetization within JPEG2000 encoding. JPEG2000, the new ISO/ITU-T standard for still image coding, has been shown to provide...
详细信息
The paper presents an efficient algorithm for post compression optimal rate allocation and packetization within JPEG2000 encoding. JPEG2000, the new ISO/ITU-T standard for still image coding, has been shown to provide superior coding efficiency to the previous standard, JPEG. However, the added efficiency of JPEG2000 comes at the cost of increased computational requirements. To improve the computational efficiency of JPEG2000, we propose a new algorithm for JPEG2000 rate allocation and packetization utilizing the D-heap data structure. Implemented in Jasper and tested on five reference images, this algorithm provides a speedup for JPEG2000's rate allocation and packetization of 15.9 times on average, and enables an average overall speedup of 33% for JPEG2000 encoding.
The detection of double JPEG compression with the same quantization matrix is a challenging problem in image forensics. In this paper, a CNN framework is proposed to solve this problem. This framework contains a prepr...
详细信息
ISBN:
(纸本)9781728102436;9789881476852
The detection of double JPEG compression with the same quantization matrix is a challenging problem in image forensics. In this paper, a CNN framework is proposed to solve this problem. This framework contains a preprocessing layer and a well-designed CNN. In the preprocessing layer, the rounding and truncation error images are extracted from continuous recompressed input samples and then fed into the following CNN. In the design of the CNN architecture, several advanced techniques are carefully considered to prevent overfitting, such as 1×1 convolutional kernel and global average pooling layer. The performance of proposed framework is evaluated on the public available image dataset (BOSSbase) with various quality factors (QF). Experimental results have shown the proposed CNN framework performs better than the state-of-the-art method based on hand-crafted features.
Key-dependent Wavelet Packet transforms (KDWPT) have been proposed for image encryption and especially for the joint application with the JPEG2000 compression framework. An assumed advantage of this compression integr...
详细信息
Key-dependent Wavelet Packet transforms (KDWPT) have been proposed for image encryption and especially for the joint application with the JPEG2000 compression framework. An assumed advantage of this compression integrated encryption scheme is its assumed negligible computational demand. In this work we analyze the assumption (KDWPT are lightweight compared to conventional encryption) both practically by experiments with state-of-the-art implementations and theoretically (by developing a proper model for the complexity of KDWPT).
Application development projects focused on the use of the JPEG/DCT data compression routine to facilitate near-real-time SAR data delivery. The key requirements were to maintain image interpretability after image com...
详细信息
Application development projects focused on the use of the JPEG/DCT data compression routine to facilitate near-real-time SAR data delivery. The key requirements were to maintain image interpretability after image compression for three distinct cases: (1) points-ship detection; (2) linear features - oil slick detection; and (3) areas-ice mapping. Maximum compression ratios and corresponding mean square-errors obtained were: ships-(53.7:1, 22.9); oil slicks (52.5:1, 25.9); ice (27.9:1, 31.8). Image interpretability was high even at the largest compression ratios for ship detection, but image interpretability degraded with increasing compression ratio for oil slick detection and ice mapping. The decrease in interpretability was due to the high frequency spatial information of the oil slick and ice mapping imagery.
暂无评论