Price Decision Support System Security – Features and Online Prediction Defense in Adversarial Environment TEO ETEROVIC 1 , DZENANA DONKO 2 Faculty of Electrical Engineering University of Sarajevo Zmaja od Bosne bb, Kampus Univerziteta, 71000 Sarajevo BOSNIA AND HERZEGOVINA 1 teterovic@etf.unsa.ba, 2 ddonko@etf.unsa.ba Abstract – Price decision support systems (PDSS) are crucial for every big retailer in order to be able to decide about product prices in hundreds of stores and thousands of products. In this paper we identify, describe and formalize several price decision support system features that can be used as an input for machine learning algorithms, after that we select the features that can be exploited by potential attackers and discuss/evaluate the security issues of online learning features in adversarial environment PDSS. At the end we propose a kernel learning defense model for the sensitive features. Keywords – Adversarial learning security, machine learning security, price decision support system security 1 Introduction Nowadays, the retailers have to be very careful and take care of the retail product prices (from now on we will use the word price) because of the high competitive surrounding. Today, more than ever, the financial success depends on price strategy of the company as the customers demand for fair prices and have a better and more up-to date review of the competitor prices. The big retail stores Walmart, BestBuy, Lidl, Aldi relay on those systems for years[4] and the PDSS is bound to their “best price” company strategy. In the 70s [1][2] the first generation of software aided price support systems appeared (“Price Maintenance”) that added standard markup to basket costs. In the 90s the first systems appeared based on Rule-based systems where software replaced existing retailer policies/rules but without any business intelligence (example: if (winter) discount all ice cream products by 20%; set all prices to end with a nine). They were very useful but the problem was they were too general, they didn’t generate any new knowledge so they only automatized the retail managers work and had no analytical capabilities to (kind of) replace the retail manager. At least in the 2000th first large chain retailers and discounters start to develop and adapt price optimization software based on machine learning. Somewhere around 2000 the first commercial price optimization software products appear on the market. An outdated list of commercial price decision support software from 2003 – mostly Rule- based (we haven’t found any open source implementations) can be found in [3], because of the evaluation and later reference we update the list with SAP® Price Optimization ex Khimetrics, SAS® Regular Price Optimization ex MarketMax, Orcle Retail Regular Price Optimization ex ProfitLogic, i2, Acnielsen, Demandtec, Evant, JDC ex Manguistics, Revionics price optimization(that where all that we could find). Still, until today, retailers (that use any price support software) mostly use Rule-based systems [1] because they are easier to adapt and more intuitional to use. But the price optimization systems based on machine learning provide the same features as rule-based systems but with heuristics and optimizations, giving the retail managers the opportunity to fit their prices politics to business politics. Price optimization systems consider price, cost and promotion history data as well as price rules and competitors data to generate prices. But still there are many( mostly small) retailers that relay on the simple cost-plus model that follows the competition prices or the supplier given prices. This can be true for small retails but for large store chains (1000+ stores) it’s very difficult(BestBuy used the word impossible) to Latest Trends in Applied Informatics and Computing ISBN: 978-1-61804-130-2 184