Doordash is one of the largest platform in the world to connect millions of local business with customers. We use advanced machine learning technologies to build a personalized customer experience and help customers discover a variant of local businesses they love. In this talk, we will introduce a few technologies we used to build our personalized homepage experience and the lessons learned during the process. Customers use our platform in different ways, they can browse on homepage, search on search bar or respond to a push notification or an email sent to them. There are also different types of actions they can take during their shopping journeys, included but not limited to views, (good) clicks, add-to-cart, and checkout. We will first introduce how we leverage customers various action sequence and transformer to build our user interest model to understand customer interests. Doordash homepage has a very vivid design containing different components and complex layout to serve our customers. The stores are organized with themes into an UI component that we call carousel. The stores, carousels and other UI components are mixed on our homepage to showcase a diverse set of options and deals customers can choose from. The complex homepage design poses challenges for homepage ranking. We build a heterogeneous ranking system to rank different type of components in a 2-D layout. Traditionally, our ranking model is optimized for conversion. However, as our business grows, we have multiple business objectives to care about. In the meanwhile, we also want to optimize for customers long term satisfaction so we can sustain and grow our platform. We will describe how do we model customers long term value and build a multi-objective ranking and optimization system to optimize and balance multiple business objectives.

Lost audio packet steganography (LACK) is a steganography technique established on the VoIP network. LACK provides a high-capacity covert channel over VoIP network by artificially delaying and dropping a number of packets in use to convey stegnogram. However, the increasing loss of packets will hurt the quality of the VoIP service. The quality deterioration will not only affect the legitimate VoIP service but also constrain the capacity of the covert channel. Discrete spring transform (DST) is proven to be a way to eliminate the perceptual redundancy in the multimedia signal. In this paper, the DST is applied on the LACK so that the perceptual redundancy of the voice frames is suppressed. In this way, the less redundant VoIP frames with perceptual equivalent quality can be transmitted in a channel whose capacity is squeezed by the established covert channel. As a result, the VoIP perceptual quality can be maintained with the existence of the covert channel. Meanwhile, the proposed DST-based method demonstrates the possibilities in exploiting the perceptual space of the multimedia signal. The simulation results show that the DST on LACK achieves up to 24 % more capacity over the LACK scheme.

Digital audio signal provides a large capacity for embedding hidden messages using digital steganography techniques. How to prevent hazardous steganography embedding on the Internet becomes an important task in the field of network security. For the Internet environment, the steganography attack method is required to be generic and real time. An active warden-based attack method is potential to be a generic method for the steganography attack. A discrete spring transform (DST)-based generic active warden steganography attack framework has been proposed by us. In this paper, based on the DST, a real-time steganography attack method is proposed. The potential unauthorized hidden message is removed in a real-time manner when uploading or downloading the audio signal. The real-time signal perceptual quality control is achieved by the automatic feedback from the objective audio quality evaluation model. The attack parameters are adaptively changed to reach a balance between the attack performance and the audio signal quality. The simulation results validate the proposed method in terms of the steganography attack performance and the audio signal quality after the attack. Copyright © 2014 John Wiley & Sons, Ltd.

There have been many robust image steganography methods that are invented in recent decades. However, this technique also may be used by malicious users to transmit dangerous information through the Internet beyond the control of the security agencies. How to detect and/or block potentially dangerous information transmission on the Internet through billions of images while not affecting the normal digital images becomes a challenging problem. Existing steganalysis methods or steganography attacking methods cannot be used for analyzing a large volume of digital images in a short time. In this paper, we propose an effective steganography attacking method that is not limited by the types of the steganography method. The proposed method can process the digital images to remove the potentially dangerous hidden information while keeping the digital image in a high visual quality. Inspired by the way in which printers and scanners work, our method decouples the human visual perception and the digital image's numerical values. It causes the numerical values of the image to change dramatically, and then the hidden information are largely damaged, while at the same time, the visual image quality can be maintained. Our experimental results have demonstrated that the peak signal-to-noise ratio of images can be above 32dB while the stego data are destroyed. Copyright © 2016 John Wiley & Sons, Ltd.

In order to prevent harmful secret information sharing by steganography, a new active-warden countermeasure approach against steganography is proposed in this paper. Differently from the other countermeasure approaches presented in current literature which need to have some knowledge of the steganographic algorithms to be attacked, our proposed method is a generic method which is independent of any particular steganography methods being utilized. In other words, our approach blindly attacks the steganography without any prior knowledge of the used algorithms or their existence. In general, the hidden information is embedded in a carrier by adjusting the coefficients of the audio or image. In our method, by exploiting the large margin between the numerical value and visual perception of the images, large amounts of visually non-detectable distortions are incurred in the image. As a result, the hidden message is destroyed by this method while the perceptual quality of the image is maintained. Inspired by the print-scan process in which most of the steganographic methods cannot survive, a transform called Discrete Spring Transform (DST) is proposed in this paper as the foundation of our attack algorithm. An image geometrization method is also developed to reconstruct the image in this paper. The simulation results have demonstrated that the PSNR of the attacked image is above 30dB with a high perceptive quality while the BER of the hidden steganographic message in the attacked image is above 0.5.

Digital steganography is often divided into two categories of research: steganographer and attacker, where a steganographer attempts to successfully hide the existence of hidden data and an attacker attempts to uncover and destroy said data. Often steganographic attacks utilize passive techniques, where the attacker monitors messages and alters or destroys the cover media only if the media is suspected to contain a message. Passive attacks are much more prevalent than active attacks for a variety of reasons. Generally passive attacks are more efficient and effective in determining the existence of hidden data in a given cover media, however, most passive attacks are not generic enough to detect the presence of a hidden message outside of a very small subset of steganographic algorithms. Conversely, current active attacks are very cost effective and often severely degrade or destroy the quality of the cover media. In this paper, we propose a novel active steganographic attack called the Discrete Spring Transform (DST). The concept of the DST attack is that of likening the cover media to a spring, in that it can be physically altered or manipulated in a manner of ways, while the basic structure and integrity of the spring is preserved despite these alterations. Since many steganographic schemes rely on the fact that the cover media remains somewhat stable the DST is effective in destroying the stego media for such types of steganographic algorithms. It follows that the DST is an effective and highly adaptable active attack that can be applied to a variety of cover media, including video. We will demonstrate that the attack is capable of defeating next-generation steganographic algorithms, including motion-vector steganography and RST-resilient algorithms by increasing the Bit Error Rate (BER) of the steganographic algorithm to approximately 0.5 while maintaining the quality of the cover media, where the Peak Signal-to-Noise Ratio (PSNR) of the image-derived media always remains above 30db.

The significant development of cell phone and mobile network in recent years leads to some security concerns caused by increasing capacity of the networks and the large margin of the multimedia data. Audio steganography is a proven way to hide information in digital audio signals, where many schemes have been effective at hiding large amounts of data. In order to prevent the malicious threat of the audio steganog-raphy, the countermeasure of audio steganography is also an important topic. In this paper, we propose a novel steganographic attack which can defeat audio steganography algorithms while maintaining an acceptable audio distortion level. The attacking method is based on a proposed transform called discrete spring transform. Similar to the time scale modification, the spring transform disables the synchronization of the hidden information. Furthermore, the proposed method has some advantages over the traditional time scale modification, therefore the steganography method which can resist to the time scale modification still can be defeated by the proposed method. The simulation results show that the proposed method successfully defeat some audio steganography methods which are resistant to the time scale modification with subjective auditory quality score no worse than -0.1.