260 likes | 459 Views
Content-Dependent Watermarking Scheme in Compressed Speech With Identifying Manner and Location of Attacks. Oscal T.-C. Chen and Chia-Hsiung Liu. Date : 2009/12/01 Speaker : Jhang Hon-Bin. Outline. Introduction Proposed Scheme Computer Simulations Conclusions. Oscal T.-C. Chen.
E N D
Content-Dependent Watermarking Scheme in Compressed Speech With Identifying Manner and Location of Attacks Oscal T.-C. Chen and Chia-Hsiung Liu Date:2009/12/01 Speaker:Jhang Hon-Bin
Outline • Introduction • Proposed Scheme • Computer Simulations • Conclusions
Oscal T.-C. Chen 美國南加州大學電機工程博士 國立中正大學電機工程學系專任教授 Chia-Hsiung Liu 國立中正大學電機工程研究所博士 Received May 31, 2006 Revised January 15, 2007 This work was supported in part by the National Science Council(NSC), Taiwan, R.O.C.
無 浮水印 加入 浮水印 發現攻擊!! Introduction (1/5) Hacker
Introduction (2/5) • Watermarking Scheme • Robust、Fragile、Semi-Fragile • Theoretical Foundation of Conventional Watermarks 固定值: Image ID 目前區塊 : Image feature 變異值: block index
Introduction (3/5) • Theoretical Foundation of Conventional Watermarks (Cont.) 鄰近區塊 : Imagefeature 識別標誌訊息 鄰近區塊 : ImageWatermark
A speech watermarking scheme must reliably determine • where and how attacks take place. • A deletion and insertion attacks that influence speech • length. • When a watermarking scheme applies (4) to verify speech • integrity. Cannot localize attacks due to an inconsistency • between Introduction (4/5) • Design Concept (Problem)
Introduction (5/5) • Objective • 準確地檢測攻擊的 位置 和攻擊的 型態 • 語音壓縮技術使用 G.723.1 • 嵌入浮水印後能保持 語音品質 最小下降
限制 的範圍 Proposed Scheme (1/8) • Design Concept (Solve) • When a watermarking scheme applies (4) to verify speech • integrity. Cannot localize attacks due to an inconsistency • between • The first scheme is to constrain the value using the • following equation:
Proposed Scheme (2/8) • Design Concept (Solve)(Count.) • The second scheme,the generating process does not use • data, such as frame index, to generate the watermark • The second approach is a watermarking scheme expressed as.
Group Group Group Frame Frame ‧ ‧ ‧ Frame Frame ‧ ‧ ‧ Frame Frame ‧ ‧ ‧ Proposed Scheme (3/8) • Design Concept (Solve)(Count.) • The two schemes can be integrated into one scheme to • help achieve our goal. • Determine attack type and localize attacked segments. Then (8) is used to generate watermarks. In most circumstances, (9) is employed to generate watermarks.
Proposed Scheme (4/8) • Watermark Generation Group g-1 Group g Last group 1 1 ‧ ‧ ‧ 1 2 ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧
由 所產生 是表示group index的位元數 Group g-1 Group g Last group ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ Proposed Scheme (5/8) • Watermark Generation (Count.) • The first frame in a group 下一個frame的pitch 目前frame的LSF 前一個group的最後一個frame
是最後一個group的frame數 是表示pitch feature的位元數 Group g-1 Group g Last group ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ Proposed Scheme (6/8) • Watermark Generation (Count.) • The lastframe ofall speech data 前一個frame 目前frame的LSF
Group g-1 Group g Last group ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ ‧ Proposed Scheme (7/8) • Watermark Generation (Count.) • The otherframe in a group 前一個frame 目前frame的LSF 下一個frame的pitch
原浮水印 Correct Or Incorrect Comparator 取出的浮水印 Proposed Scheme (8/8) 產生浮水印 & 嵌入浮水印 1 2 3 ITUT G.723.1 MP-MLQ (6.3 k bit/s) 1 frame= 240 取出浮水印 60 60 60 60 04 / 23 / 2009 P16
Computer Simulations (1/8) • Evaluation of Speech Quality • Eleven conversations of 60.6 ~125.6 s are encoded. 表3.1 :正常語音壓縮和加入浮水印後PESQ數值
Computer Simulations (2/8) • Comparison Between Proposed and Conventional Schemes • The schemes proposed by Yuan et al. and Steinebach et al. • are rederived from (5)–(7) to (15)–(17), respectively
Computer Simulations (3/8) 表3.2 : 三種浮水印攻擊對原 2019 frames語音的影響
Computer Simulations (4/8) • The Watermarking Scheme Based on (15)
Computer Simulations (5/8) • The Watermarking Scheme Based on(16)
Computer Simulations (6/8) • The Watermarking Scheme Based on(17)
Computer Simulations (7/8) • The Proposed Scheme
Computer Simulations (8/8) 表3.3 : Performance Comparison of The Proposed and Conventional Schemes
Conclusions (1/1) • This scheme could locate counterfeit frames and • identify the manner of counterfeiting the data. • Only decreases the PESQ by 0.11