《计算机应用》唯一官方网站 ›› 2024, Vol. 44 ›› Issue (11): 3479-3486.DOI: 10.11772/j.issn.1001-9081.2023101518
林翔1,2, 金彪1,2, 尤玮婧1,2, 姚志强1,2, 熊金波1,2()
收稿日期:
2023-11-07
修回日期:
2024-01-10
接受日期:
2024-01-12
发布日期:
2024-11-13
出版日期:
2024-11-10
通讯作者:
熊金波
作者简介:
林翔(1996—),男,福建厦门人,硕士研究生,CCF会员,主要研究方向:人工智能安全基金资助:
Xiang LIN1,2, Biao JIN1,2, Weijing YOU1,2, Zhiqiang YAO1,2, Jinbo XIONG1,2()
Received:
2023-11-07
Revised:
2024-01-10
Accepted:
2024-01-12
Online:
2024-11-13
Published:
2024-11-10
Contact:
Jinbo XIONG
About author:
LIN Xiang, born in 1996, M. S. candidate. His research interests include artificial intelligence security.Supported by:
摘要:
预训练模型容易受到外部敌手实施的模型微调和模型剪枝等攻击,导致它的完整性被破坏。针对这一问题,提出一种针对黑盒模型的脆弱指纹框架FFWAS (Fragile Fingerprint With Adversarial Samples)。首先,提出一种无先验知识的模型复制框架,而FFWAS为每一位用户创建独立的模型副本;其次,利用黑盒方法在模型边界放置脆弱指纹触发集,若模型发生修改,边界发生变化,触发集将被错误分类;最后,用户借助模型副本上的脆弱指纹触发集对模型的完整性进行验证,若触发集的识别率低于预设阈值,则意味着模型完整性已被破坏。基于2种公开数据集MNIST和CIFAR-10对FFWAS的有效性和脆弱性进行实验分析,结果表明,在模型微调和剪枝攻击下,FFWAS的指纹识别率相较于完整模型均明显下降并低于设定阈值;与基于模型唯一性和脆弱签名的深度神经网络认证框架(DeepAuth)相比,FFWAS的触发集与原始样本在2个数据集上的相似性分别提高了约22%和16%,表明FFWAS具有更好的隐蔽性。
中图分类号:
林翔, 金彪, 尤玮婧, 姚志强, 熊金波. 基于脆弱指纹的深度神经网络模型完整性验证框架[J]. 计算机应用, 2024, 44(11): 3479-3486.
Xiang LIN, Biao JIN, Weijing YOU, Zhiqiang YAO, Jinbo XIONG. Model integrity verification framework of deep neural network based on fragile fingerprint[J]. Journal of Computer Applications, 2024, 44(11): 3479-3486.
数据集 | 模型 | 模型精度 | FSR |
---|---|---|---|
MNIST | LeNet-5 | 98.86 | N/A |
模型副本 | 98.11 | N/A | |
指纹模型副本 | 98.11 | 100 | |
CIFAR-10 | ResNet-34 | 95.09 | N/A |
模型副本 | 93.80 | N/A | |
指纹模型副本 | 93.80 | 100 |
Tab.1 Comparison of precision and FSR of original model, model copy and fingerprint model copy on two datasets
数据集 | 模型 | 模型精度 | FSR |
---|---|---|---|
MNIST | LeNet-5 | 98.86 | N/A |
模型副本 | 98.11 | N/A | |
指纹模型副本 | 98.11 | 100 | |
CIFAR-10 | ResNet-34 | 95.09 | N/A |
模型副本 | 93.80 | N/A | |
指纹模型副本 | 93.80 | 100 |
数据集 | 模型 | 模型精度 | FSR |
---|---|---|---|
MNIST | 指纹LeNet-5 | 98.11 | 100 |
微调LeNet-5 | 99.02 | 14 | |
CIFAR-10 | 指纹ResNet-34 | 93.80 | 100 |
微调ResNet-34 | 93.53 | 2 |
表2 FFWAS在模型微调下的FSR与模型精度对比 ( %)
Tab.2 Comparison of FSR and model precision of FFWAS under model fine-tuning
数据集 | 模型 | 模型精度 | FSR |
---|---|---|---|
MNIST | 指纹LeNet-5 | 98.11 | 100 |
微调LeNet-5 | 99.02 | 14 | |
CIFAR-10 | 指纹ResNet-34 | 93.80 | 100 |
微调ResNet-34 | 93.53 | 2 |
数据集 | 文献[ | DeepAuth[ | 文献[ | 文献[ | FFWAS |
---|---|---|---|---|---|
MNIST | 3.9×10-3 | 4.1×10-4 | — | — | 3.2×10-4 |
CIFAR-10 | 2.4×10-4 | 6.9×10-5 | 352.37 | 455.78 | 5.8×10-5 |
表3 原始样本与不同触发集之间的平均L2距离比较
Tab. 3 Comparison of average L2 distance between original samples and different trigger sets
数据集 | 文献[ | DeepAuth[ | 文献[ | 文献[ | FFWAS |
---|---|---|---|---|---|
MNIST | 3.9×10-3 | 4.1×10-4 | — | — | 3.2×10-4 |
CIFAR-10 | 2.4×10-4 | 6.9×10-5 | 352.37 | 455.78 | 5.8×10-5 |
数据集 | 文献[ | DeepAuth[ | 文献[ | 文献[ | FFWAS |
---|---|---|---|---|---|
MNIST | 0.08 | 12.62 | — | — | 5.29 |
CIFAR-10 | 0.13 | 30.48 | 0 | 26.58 | 17.61 |
表4 不同触发集生成每张图片的平均时间比较 ( s)
Tab.4 Comparison of average generation time for each image of different trigger sets
数据集 | 文献[ | DeepAuth[ | 文献[ | 文献[ | FFWAS |
---|---|---|---|---|---|
MNIST | 0.08 | 12.62 | — | — | 5.29 |
CIFAR-10 | 0.13 | 30.48 | 0 | 26.58 | 17.61 |
方案 | 方法类型 | 触发集生成 | 验证方法 | |
---|---|---|---|---|
文献[ | 鲁棒水印 | 黑盒 | 黑盒 | -1.0 |
文献[ | 鲁棒指纹 | 白盒 | 黑盒 | -0.7 |
DeepAuth[ | 脆弱水印 | 白盒 | 黑盒 | -0.3 |
文献[ | 脆弱水印 | 黑盒 | 黑盒 | -0.1 |
FFWAS | 脆弱指纹 | 黑盒 | 黑盒 | = |
表5 不同DNN水印方案的比较
Tab.5 Comparison of different DNN watermarking frameworks
方案 | 方法类型 | 触发集生成 | 验证方法 | |
---|---|---|---|---|
文献[ | 鲁棒水印 | 黑盒 | 黑盒 | -1.0 |
文献[ | 鲁棒指纹 | 白盒 | 黑盒 | -0.7 |
DeepAuth[ | 脆弱水印 | 白盒 | 黑盒 | -0.3 |
文献[ | 脆弱水印 | 黑盒 | 黑盒 | -0.1 |
FFWAS | 脆弱指纹 | 黑盒 | 黑盒 | = |
1 | HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2016: 770-778. |
2 | AMODEI D, ANANTHANARAYANAN S, ANUBHAI R, et al. Deep Speech 2: end-to-end speech recognition in English and Mandarin[C]// Proceedings of the 33rd International Conference on Machine Learning. Cambridge: MIT Press, 2016: 173-182. |
3 | 田蕾,葛丽娜.基于差分隐私的广告推荐算法[J].计算机应用,2023,43(11):3346-3350. |
TIAN L, GE L N. Advertising recommendation algorithm based on differential privacy[J]. Journal of Computer Applications, 2023, 43(11): 3346-3350. | |
4 | RIBEIRO M, GROLINGER K, CAPRETZ M A M. MLaaS: machine learning as a service[C]// Proceedings of the IEEE 14th International Conference on Machine Learning and Applications. Piscataway: IEEE, 2015: 896-902. |
5 | VAN SCHYNDEL R G, TIRKEL A Z, OSBORNE C F. A digital watermark[C]// Proceedings of 1st International Conference on Image Processing: Volume 2. Piscataway: IEEE, 1994: 86-90. |
6 | ZHANG X, WANG S. Fragile watermarking with error-free restoration capability[J]. IEEE Transactions on Multimedia, 2008, 10(8): 1490-1499. |
7 | UCHIDA Y, NAGAI Y, SAKAZAWA S, et al. Embedding watermarks into deep neural networks[C]// Proceedings of the 2017 ACM International Conference on Multimedia Retrieval. New York: ACM, 2017: 269-277. |
8 | CHEN H, ROUHANI B D, FU C, et al. DeepMarks: a secure fingerprinting framework for digital rights management of deep learning models[C]// Proceedings of the 2019 International Conference on Multimedia Retrieval. New York: ACM, 2019: 105-113. |
9 | ZHANG J, GU Z, JANG J, et al. Protecting intellectual property of deep neural networks with watermarking[C]// Proceedings of the 2018 ACM Asia Conference on Computer and Communications Security. New York: ACM, 2018: 159-172. |
10 | LI Z, HU C, ZHANG Y, et al. How to prove your model belongs to you: a blind-watermark based framework to protect intellectual property of DNN[C]// Proceedings of the 35th Annual Computer Security Applications Conference. New York: ACM, 2019: 126-137. |
11 | ADI Y, BAUM C, CISSE M, et al. Turning your weakness into a strength: watermarking deep neural networks by backdooring[C]// Proceedings of the 27th USENIX Security Symposium. Berkeley: USENIX Association, 2018: 1615-1631. |
12 | JIA H, CHOQUETTE-CHOO C A, CHANDRASEKARAN V, et al. Entangled watermarks as a defense against model extraction[C]// Proceedings of the 30th USENIX Security Symposium. Berkeley: USENIX Association, 2021: 1937-1954. |
13 | 樊雪峰,周晓谊,朱冰冰,等. 深度神经网络模型版权保护方案综述[J]. 计算机研究与发展, 2022, 59(5): 953-977. |
FAN X F, ZHOU X Y, ZHU B B, al at. Survey of copyright protection schemes based on DNN model[J]. Journal of Computer Research and Development, 2022, 59(5): 953-977. | |
14 | 李璇,邓天鹏,熊金波,等.基于模型后门的联邦学习水印[J].软件学报,2024,35(7):3454-3468. |
LI X, DENG T P, XIONG J B, et al. Federated learning watermark based on backdoor[J]. Journal of Software, 2024, 35(7): 3454-3468. | |
15 | WAGNER N R. Fingerprinting[C]// Proceedings of the 1983 IEEE Symposium on Security and Privacy. Piscataway: IEEE, 1983: 18. |
16 | ZHAO J, HU Q, LIU G, et al. AFA: adversarial fingerprinting authentication for deep neural networks[J]. Computer Communications, 2020, 150: 488-497. |
17 | XUE M, WU Z, HE C, et al. Active DNN IP protection: a novel user fingerprint management and DNN authorization control technique[C]// Proceedings of the IEEE 19th International Conference on Trust, Security and Privacy in Computing and Communications. Piscataway: IEEE, 2020: 975-982. |
18 | XUE M, HE C, WANG J, et al. One-to-N & N-to-One: two advanced backdoor attacks against deep learning models[J]. IEEE Transactions on Dependable and Secure Computing, 2022, 19(3): 1562-1578. |
19 | XUE M, SUN S, ZHANG Y, et al. Active intellectual property protection for deep neural networks through stealthy backdoor and users' identities authentication[J]. Applied Intelligence, 2022, 52(14): 16497-16511. |
20 | GUAN X, FENG H, ZHANG W, et al. Reversible watermarking in deep convolutional neural networks for integrity authentication[C]// Proceedings of the 28th ACM International Conference on Multimedia. New York: ACM, 2020: 2273-2280. |
21 | BOTTA M, CAVAGNINO D, ESPOSITO R. NeuNAC: a novel fragile watermarking algorithm for integrity protection of neural networks[J]. Information Sciences, 2021, 576: 228-241. |
22 | HE Z, ZHANG T, LEE R. Sensitive-sample fingerprinting of deep neural networks[C]// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2019: 4724-4732. |
23 | LAO Y, ZHAO W, YANG P, et al. DeepAuth: a DNN authentication framework by model-unique and fragile signature embedding[C]// Proceedings of the 36th AAAI Conference on Artificial Intelligence. Palo Alto, CA: AAAI Press, 2022: 9595-9603. |
24 | MĄDRY A, MAKELOV A, SCHMIDT L, et al. Towards deep learning models resistant to adversarial attacks[EB/OL]. [2023-08-08]. . |
25 | WANG S, ABUADBBA S, AGARWAL S, et al. PublicCheck: public integrity verification for services of run-time deep models[C]// Proceedings of the 2023 IEEE Symposium on Security and Privacy. Piscataway: IEEE, 2023: 1348-1365. |
26 | CARLINI N, WAGNER D. Towards evaluating the robustness of neural networks[C]// Proceedings of the 2017 IEEE Symposium on Security and Privacy. Piscataway: IEEE, 2017: 39-57. |
27 | ZHU R, WEI P, LI S, et al. Fragile neural network watermarking with trigger image set[C]// Proceedings of the 2021 International Conference on Knowledge Science, Engineering and Management, LNCS 12815. Cham: Springer, 2021: 280-293. |
28 | HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network[EB/OL]. [2023-10-22]. . |
29 | LeCUN Y, BOTTOU L, BENGIO Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324. |
30 | KRIZHEVSKY A. Learning multiple layers of features from tiny images[R/OL]. [2023-07-19]. . |
31 | KINGMA D P, BA J L. Adam: a method for stochastic optimization[EB/OL]. [2022-12-22].. |
32 | HOOKER S, DAUPHIN Y, COURVILLE A, et al. Selective brain damage: measuring the disparate impact of model pruning[EB/OL]. [2023-09-26]. . |
33 | HAN S, POOL J, TRAN J, et al. Learning both weights and connections for efficient neural network[C]// Proceedings of the 28th International Conference on Neural Information Processing Systems — Volume 1. Cambridge: MIT Press, 2015: 1135-1143. |
34 | PITTARAS N, MARKATOPOULOU F, MEZARIS V, et al. Comparison of fine-tuning and extension strategies for deep convolutional neural networks[C]// Proceedings of the 2017 International Conference on MultiMedia Modeling, LNCS 10132. Cham: Springer, 2017: 102-114. |
[1] | 杨兴耀, 陈羽, 于炯, 张祖莲, 陈嘉颖, 王东晓. 结合自我特征和对比学习的推荐模型[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2704-2710. |
[2] | 杜郁, 朱焱. 构建预训练动态图神经网络预测学术合作行为消失[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2726-2731. |
[3] | 王娜, 蒋林, 李远成, 朱筠. 基于图形重写和融合探索的张量虚拟机算符融合优化[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2802-2809. |
[4] | 李云, 王富铕, 井佩光, 王粟, 肖澳. 基于不确定度感知的帧关联短视频事件检测方法[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2903-2910. |
[5] | 唐廷杰, 黄佳进, 秦进. 基于图辅助学习的会话推荐[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2711-2718. |
[6] | 张睿, 张鹏云, 高美蓉. 自优化双模态多通路非深度前庭神经鞘瘤识别模型[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2975-2982. |
[7] | 秦璟, 秦志光, 李发礼, 彭悦恒. 基于概率稀疏自注意力神经网络的重性抑郁疾患诊断[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2970-2974. |
[8] | 方介泼, 陶重犇. 应对零日攻击的混合车联网入侵检测系统[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2763-2769. |
[9] | 杨航, 李汪根, 张根生, 王志格, 开新. 基于图神经网络的多层信息交互融合算法用于会话推荐[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2719-2725. |
[10] | 姚光磊, 熊菊霞, 杨国武. 基于神经网络优化的花朵授粉算法[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2829-2837. |
[11] | 黄颖, 杨佳宇, 金家昊, 万邦睿. 用于RGBT跟踪的孪生混合信息融合算法[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2878-2885. |
[12] | 陈虹, 齐兵, 金海波, 武聪, 张立昂. 融合1D-CNN与BiGRU的类不平衡流量异常检测[J]. 《计算机应用》唯一官方网站, 2024, 44(8): 2493-2499. |
[13] | 杨莹, 郝晓燕, 于丹, 马垚, 陈永乐. 面向图神经网络模型提取攻击的图数据生成方法[J]. 《计算机应用》唯一官方网站, 2024, 44(8): 2483-2492. |
[14] | 赵宇博, 张丽萍, 闫盛, 侯敏, 高茂. 基于改进分段卷积神经网络和知识蒸馏的学科知识实体间关系抽取[J]. 《计算机应用》唯一官方网站, 2024, 44(8): 2421-2429. |
[15] | 张春雪, 仇丽青, 孙承爱, 荆彩霞. 基于两阶段动态兴趣识别的购买行为预测模型[J]. 《计算机应用》唯一官方网站, 2024, 44(8): 2365-2371. |
阅读次数 | ||||||
全文 |
|
|||||
摘要 |
|
|||||