Michiel van der Veen

Audio watermarking for monitoring and copy protection

Based on existing technology used in image and video watermarking, we have developed a robust audio watermarking technique. The embedding algorithm operates in frequency domain, where the magnitudes of the Fourier coefficients are slightly modified. In the temporal domain, an additional scale parameter and gain function are necessary to refine the watermark and achieve perceptual transparency. Watermark detection relies on the Symmetrical Phase Only Matched Filtering (SPOMF) cross-correlation approach. Not only the presence of a watermark, but also its cyclic shift is detected. This shift supports a multi-bit payload for one particular watermark sequence. The watermarking technology proved to be very robust to a large number of signal processing “attacks” such as MP3 (64 kb/s), all-pass filtering, echo addition, time-scale modification, resampling, noise addition, etc. It is expected that this approach may contribute in a wide variety of existing (e.g. monitoring and copy protection) and future applications.

Secure watermark embedding through partial encryption

Secure watermark embedding allows to securely embed a watermark into a piece of content at an untrusted user device without compromising the security of the watermark key, the watermark or the original. In this paper, we show how secure embedding can be achieved by using traditional watermarking schemes in conjunction with partial encryption techniques, which were primarily developed to facilitate fast encryption of media content. Based on this concept, we develop two new efficient secure embedding mechanisms, one for the MASK watermarking scheme operating on baseband audio and one for a spread spectrum watermarking scheme operating on MPEG-2 encoded video streams.

High capacity reversible watermarking for audio

A digital watermark can be seen as an information channel, which is hidden in a cover signal. It is usually designed to be imperceptible to human observers. Although imperceptibility is often achieved, the inherent modification of the cover signal may be viewed as a potential disadvantage. In this paper, we present a reversible watermarking technique for digital audio signals. In our context reversibility refers to the ability to restore the original input signal in the watermark detector. In summary, the approach works as follows. In the encoder, the dynamic range of the input signal is limited (i.e. it is compressed), and part of the unused bits is deployed for encoding the watermark bits. Another part of these bits is used to convey information for the bit-exact reconstruction of the cover signal. It is the purpose of the watermark detector to extract the watermark and reconstruct the input signal by restoring the original dynamic range. In this study we extensively tested this new algorithm with a variety of settings using audio items with different characteristics. These experiments showed that for 16bit PCM audio, capacities close to 1-bit per sample can be achieved, while perceptual degradation of the watermarked signal remained acceptable.

Face Biometrics with Renewable Templates

In recent literature, privacy protection technologies for biometric templates were proposed. Among these is the so-called helper-data system (HDS) based on reliable component selection. In this paper we integrate this approach with face biometrics such that we achieve a system in which the templates are privacy protected, and multiple templates can be derived from the same facial image for the purpose of template renewability. Extracting binary feature vectors forms an essential step in this process. Using the FERET and Caltech databases, we show that this quantization step does not significantly degrade the classification performance compared to, for example, traditional correlation-based classifiers. The binary feature vectors are integrated in the HDS leading to a privacy protected facial recognition algorithm with acceptable FAR and FRR, provided that the intra-class variation is sufficiently small. This suggests that a controlled enrollment procedure with a sufficient number of enrollment measurements is required.

Informed detection of audio watermark for resolving playback speed modifications

In this paper, we present a method for informed watermark detection in audio signals. In this context, informed detection refers to the mechanism in which the original media signal is used to improve robustness and/or complexity. Here, we focused on reducing complexity by addressing the geometrical distortion problem. In audio, geometrical distortion translates to time scaling. Generally, s speed changes of the audio signals necessitate a mechanism in the detector to retrieve the watermark. Usually this is done by some kind of (clever) exhaustive search algorithms. In this paper we show that the speed searching process can significantly be improved using fingerprinting technology. Particularly, an initial estimate of the speed is obtained by extracting the so-called local fingerprint time stamps. This initial estimate is then used as a starting point for the exhaustive speed search algorithm in the watermark detector. We also give analyses on the accuracies of the watermark detector and the fingerprint algorithm and the relations thereof. Moreover, we present the concept for integrating fingerprint algorithm in the watermark detector and discuss the influence of various parameters. We show that, apart from a significant complexity reduction, such integration can lead to an improvement in robustness as a byproduct.

Reversible watermarking for images

Reversible watermarking is a technique for embedding data in a digital host signal in such a manner that the original host signal can be restored in a bit-exact manner in the restoration process. In this paper, we present a general framework for reversible watermarking in multi-media signals. A mapping function, which is in general neither injective nor surjective, is used to map the input signal to a perceptually equivalent output signal. The resulting unused sample values of the output signal are used to encode additional (watermark) information and restoration data. At the 2003 SPIE conference, examples of this technique applied to digital audio were presented. In this paper we concentrate on color and gray-scale images. A particular challenge in this context is not only the optimization of rate-distortion, but also the measure of perceptual quality (i.e. the distortion). In literature distortion is often expressed in terms of PSNR, making comparison among different techniques relatively straightforward. We show that our general framework for reversible watermarking applies to digital images and that results can be presented in terms of PSNR rate-distortions. However, the framework allows for more subtle signal manipulations that are not easily expressed in terms of PSNR distortion. These changes involve manipulations of contrast and/or saturation.

Electronic content delivery and forensic watermarking

Today, digital content distribution is already common practice for media types like audio. It is expected that in the near future, systems like Video-on-Demand will also increase in popularity. Many Digital Right Management (DRM) models exist enforcing the copyrights associated with the distributed content. We argue that a simplified model based on so-called forensic watermarks is a viable alternative approach that closely mimics the current model of physical music and video distribution by providing a reasonable compromise between the conflicting interests of consumers and content owners. In this paper, we address the forensic watermark DRM model and introduce a generalized architecture for the distribution server and the forensic watermark tracker. In this model, important attack scenarios are identified as being the ‘copy attack’ and the watermark ‘rendering’. We show that it is more difficult to successfully work out these attack scenarios by (i) integrating a content identification and watermarking system and (ii) making the watermark (both the carrier signal as well as the payload) content dependent. In addition to these security aspects, we also focus on efficiency of the distribution server. By separating the computational efforts in so-called pre-coding and on-line computational phases, the required signal processing resources of the distribution server can be reduced significantly. The general architecture including security and efficiency requirements are analyzed in a practical example of Electronic Music Delivery (EMD). We show that computational complexity can be reduced up to a factor 4 compared to a straightforward approach. In practice, it means that with current state-of-the art computers (Pentium IV, 1 GHz) watermark embedding speeds of about 40 times Real Time (RT) can be achieved. In addition to the embedding architecture, we also demonstrate an efficient method for detecting watermarks in a forensic watermark tracker. Having multimedia identification fingerprints available (required for working with content-dependent watermarks), the efficiency of watermark detection can be enhanced by using them as side information. Again, in our EMD prototype, we demonstrate that (depending on the parameter settings) the detector can run up to 50 times faster than a blind detector, in which the original or the fingerprint is not available.

Reference point detection for improved fingerprint matching

One of the important stages of fingerprint recognition is the registration of the fingerprints with respect to the original template. This is not a straightforward task as fingerprint images may have been subject to rotations and translations. Popular techniques for fingerprint registration use a reference point to achieve alignment. The drawback of existing methods of core/reference point detection is their poor performance on rotated images. In this paper, we propose a new approach for rotation invariant and reliable reference point detection applicable to fingerprints of different quality and types. Our approach is based on the integration of a directional vector field (representing the doubled ridge orientations in fingerprints) over a closed contour. We define the reference point as the point of the highest curvature. Areas of high curvature in the fingerprint are characterized by large differences in the orientations and correspond to high curvatures in the directional vector fields. Closed contour integrals of orientation vector field, defined as above, over a circle centered around the reference point corresponds to maximal closed curve integrals, and the values associated with such integrals are rotation invariant. Experimental results prove that with the proposed approach we can locate the reference point with high accuracy. Comparison with existing methods is provided.

New modulation-based watermarking technique for video

Successful watermarking algorithms have already been developed for various applications ranging from meta-data tagging to forensic tracking. Nevertheless, it is commendable to develop alternative watermarking techniques that provide a broader basis for meeting emerging services, usage models and security threats. To this end, we propose a new multiplicative watermarking technique for video, which is based on the principles of our successful MASK audio watermark. Audio-MASK has embedded the watermark by modulating the short-time envelope of the audio signal and performed detection using a simple envelope detector followed by a SPOMF (symmetrical phase-only matched filter). Video-MASK takes a similar approach and modulates the image luminance envelope. In addition, it incorporates a simple model to account for the luminance sensitivity of the HVS (human visual system). Preliminary tests show algorithms transparency and robustness to lossy compression.

Multistage face recognition using adaptive feature selection and classification

In this paper, we propose a cascaded face-identification framework for enhanced recognition performance. During each stage, the classification is dynamically optimized to discriminate a set of promising candidates selected from the previous stage, thereby incrementally increasing the overall discriminating performance. To ensure improved performance, the base classifier at each stage should satisfy two key properties: (1) adaptivity to specific populations, and (2) high training and identification efficiency such that dynamic training can be performed for each test case. To this end, we adopt a base classifier with (1) dynamic person-specific feature selection, and (2) voting of an ensemble of simple classifiers based on selected features. Our experiments show that the cascaded framework effectively improves the face recognition rate by up to 5% compared to a single stage algorithm, and it is 2-3% better than established well-known face recognition algorithms.