With the advent of sensors, more and more services are developed in order to provide customers with insights about their health and their appliances’ energy consumption at home. To do so, these services use new mining algorithms that create new inference channels. However, the collected sensor data can be diverted to infer personal data that customers do not consent to share. This indirect access to data that are not collected corresponds to inference attacks involving raw sensor data (IASD). Towards these new kinds of attacks, existing inference detection systems do not suit the representation requirements of these inference channels and of user knowledge. In this paper, we propose RICE-M (Raw sensor data based Inference ChannEl Model) that meets these inference channel representations. Based on RICE-M, we proposed RICE-Sy an extensible system able to detect IASDs, and evaluated its performance taking as a case study the MHEALTH dataset. As expected, detecting IASD is proven to be quadratic due to huge sensor data managed and a quickly growing amount of user knowledge. To overcome this drawback, we propose first a set of conceptual optimizations that reduces the detection complexity. Although becoming linear, as online detection time remains greater than a fixed acceptable query response limit, we propose two approaches to estimate the potential of RICE-Sy. The first one is based on partitioning strategies which aim at partitioning the knowledge of users. We observe that by considering the quantity of knowledge gained by a user as a partitioning criterion, the median detection time of RICE-Sy is reduced by 63%. The second approach is H-RICE-SY, a hybrid detection architecture built on RICE-Sy which limits the detection at query-time to users that have a high probability to be malicious. We show the limits of processing all malicious users at query-time, without impacting the query answer time. We observe that for a ratio of 30% users considered as malicious, the median online detection time stays under the acceptable time of 80 , for up to a total volume of 1.2 million user knowledge entities. Based on the observed growth rates, we have estimated that for 5% of user knowledge issued by malicious users, a maximum volume of approximately 8.6 million user’s information can be processed online in an acceptable time.