Users in online social networks are vulnerable to attribute inference attacks due to some published data. Thus, the picture owner's gender has a strong influence on individuals' emotional reactions to the photo. In this work, we present a graph-embedding approach for gender inference attacks based on pictures meta-data such as (i) alt-texts generated by Facebook to describe the content of images, and (ii) Emojis/Emoticons posted by friends, friends of friends or regular users as a reaction to the picture. Specifically, we apply a semi-supervised technique, node2vec, for learning a mapping of pictures meta-data to a low-dimensional vector space. Next, we study in this vector space the gender closeness of users who published similar photos and/or received similar reactions. We leverage this image sharing and reaction mode of Facebook users to derive an efficient and accurate technique for user gender inference. Experimental results show that privacy attack often succeeds even when other information than pictures published by their owners is either hidden or unavailable.