Susheel Suresh

I finished my undergradute studies in Computer Science at P.E.S Institute of Technology in Aug 2017. During undergrad days, I worked as a student researcher at the Centre for Cloud Computing and Big Data (CCBD).

Currently, I work as a product intern Member of Technical Staff at Adobe Systems, Bangalore. I use NLP and Deep Learning techniques for solving digital marketing problems.

Previously, I worked as a research intern at INRIA, Paris with Prof. Catuscia Palamidessi and was a summer research intern at IIT Bombay under Prof. R.K Shyamasundar. I have also spent time as a student intern at Nokia Developer Lab, Bangalore. I am fortunate to have had the guidance of Kostas Chatzikokolakis (CNRS) and Narendra Kumar (TIFR) during my research internships.

I play carnatic classical music on the electric guitar (strat) and my guru is mandolin vidwan Prasanna Ballal who trained under mandolin maestro Late U. Srinivas.

Email  /  CV  /  LinkedIn  /  Github  /  Google Scholar


  • [Feb 2018] Attending AAAI 2018 for paper presentaion @ New Orleans, USA. Check out the slides.
  • [Nov 2017] VoC-DL paper accepted @ AAAI 2018.
  • [Aug 2017] Starting new job as Member Technical Staff @ Adobe, Bangalore.
  • [Jul 2017] Attending Deep Learning Summer School @ CVIT - IIIT Hyderabad, India.
  • [Jan 2017] Starting Internship at Adobe, Bangalore.
  • [Dec 2016] 3D Image Analysis Paper accepted @ ICVGIP 2016. Checkout the program schedule. Quick intro slides.
  • [Mar 2016] Will be at INRIA Saclay for research internship this summer.
  • [Dec 2015] Presented the paper on Secure Data Sharing @ ICISS 2015, Kolkata India. The presentation is available here. Video demo of the attack and it's mitigation is here.



My interests are primarily focused on three interleaved areas: Machine Learning, Computer Vision and Data Science. I have worked on segmentation, object detection, 3D reconstruction and application of these concepts to real world problems. In the security and privacy field, I have worked on Differential Privacy (it's application to location privacy) and Information Flow Control.

VoC-DL : Revisiting Voice of Customer using Deep Learning

Susheel Suresh, Guru Rajan T S, Vipin Gopinath

Thirty Second AAAI/IAAI - 2018 (emerging track paper), 2-7 Feb 2018

paper | abstract | presentation

In the field of digital marketing, understanding the voice of the customer is paramount. Mining textual content written by visitors on websites or social media can offer new dimensions to marketers and CX executives. Traditional tasks in NLP like sentiment analysis, topic modeling etc. can solve only certain specific problems but dont provide a generic solution to identifying/understanding the intention behind a text. In this paper we consider higher dimensional extensions to the sentiment concept by incorporating labels like product enquiry, buying intent, seeking help, feedback and pricing query which give us a deeper understanding of the text. We show how our model performs on a real world enterprise use case. Word2Vec embeddings are used for word representations and later we compare three algorithms for classification. SVM's provide us with a strong baseline. Two deep learning models viz. vanilla CNN and RNN's with LSTM are compared.With no use of hard coded or hand engineered features, our generic model can be used in a variety of use cases where text mining is involved with ease.

Towards Quantifying the Amount of Uncollected Garbage through Image Analysis

Susheel Suresh, Tarun Sharma, Prashanth T.K., Subramaniam V, Dinkar Sitaram & Nirupama M.

ICVGIP 2016, December 18-22, 2016

paper | abstract | poster | code

Civic authorities in many Indian cities have a tough time in garbage collection and as a result there is a pile up of garbage in the cities. In order to manage the situation, it is first required to be able to quantify the issue. In this paper, we address the problem of quantification of garbage in a dump using a two step approach. In the first step, we build a mobile application that allows citizens to capture images of garbage and upload them to a server. The second step, in the back-end performs analysis on these images to estimate the amount of garbage using computer vision techniques. Our approach to volume estimation uses multiple images of the same dump (provided by the mobile application) from different perspectives, segments the dump from the background, reconstructs a three dimensional view of the dump and then estimates its volume. Using our novel pipeline, our experiments indicate that with 8 different perspectives, we are able to achieve an accuracy of about 85 % for estimating the volume.

Enforcing Secure Data Sharing in Web Application Development Frameworks Like Django Through Information Flow Control

Susheel Suresh, N. V. Narendra Kumar, R. K. Shyamasundar.

Conference Version: ICISS 2015, December 16-20, 2015.
Published in LNCS Information Systems Security, Volume 9478

paper | abstract | code

The primary aim of web application development frameworks like Django is to provide a platform for developers to realize applications from concepts to launch as quickly as possible. While Django framework provides hooks that enable the developer to avoid the common security mistakes, there is no systematic way to assure compliance of a security policy while developing an application from various components. In this paper, we show the security flaws that arise by considering different versions of an application package and then show how, these mistakes that arise due to incorrect flow of information can be overcome using the Readers-Writers Flow Model that has the ability to manage the release and subsequent propagation of information.


Experiments on Location Privacy based on Differential Privacy

Susheel Suresh, Kostas Chatzikokolakis and Catuscia Palamidessi

INRIA Comete Lab, Summer 2016

abstract | code

We study the effects of an adversary using semantic information for constructing his priors and statergy, on the user's privacy. For the study, we assume that the adversary obtains the semantic meaning of the users actual location from some side channel and uses it to chart out 'informed' priors which will reduce his loss. The Side Channel in this case could be any geosocial networking application like Twitter, Facebook, Foursquare, Yelp, etc.
    The adversary makes use of this semantic meaning in- ferred from a side channel to obtain Location Types. Lo- cation Types are a class of broad labels for POIs, for eg., Restaturant, Train Station, Govt. Office, Park, Airport, Theater, etc. With this locatin type, the adversary then constructs global and user priors from the datasets. For eg., if the semantic label inferred from the side channel was 'Park', then the adversary would use this label and query all POIs from the dataset which have a label 'Park' and then construct his priors, rather than just considering all the POIs. These priors of POIs with 'Park' label will give the adversary a general sense of users possible locations in the respective region defined by the dataset. With the help of these global and user priors from location based datasets, the adversary creates an effective remapping strategy.

Real Time Video Content Based Contextual Advertisement

Susheel Suresh, Tarun Sharma

"What the Hack!" SAP Labs India Hackathon, 2015

demo video | tech report | abstract | code

With the advent of intermediary commercial ad-networks in charge of optimizing ad selection with a twin goal of increasing revenue and improving user experience, it is preferable to have ads relevant to the web content than generic ones. This method of advertising is popularly known as Contextual Advertising. Advertising has become ubiquitous in the internet community and more so in the ever-growing and popular online video delivery websites (e.g. YouTube, Vimeo). Video advertising is becoming increasingly popular on these websites as it has the most user engagement levels.
    Our novel method automatically associates ads from an advertisement database and seamlessly recommends them at the appropriate time within each individual video in real-time. If a context switch occurs in the video a different relevant ad is placed. Ads are in the form of banners (containing visuals and text) and are placed beside a playing video. Unlike most video sites which treat video advertising as general text advertising by displaying general, contextually irrelevant video ads at the beginning, end and beside a playing video, our approach aims to recommend contextually relevant ads for a video stream taking into account the audio/speech aspect. Specifically, given a Web page containing an online video, our method is able to extract keywords/text in the form of phrases from the video segment, perform entity extraction on the phrase, detect and resolve ambiguous entities and finally place relevant ads.

Pothole Detection Using Machine Learning

"Inmobi Spring HackDay" InMobi, 2016

Made a crowd sourced Android App which would record accelerometer readings and GPS values as a rider travelled. Features are then extracted using a sliding overlapping window, and passed to a SVM to predict the occurrence of potholes or accidents. Using previous data, potholes are ranked in order of danger posed and this data is visualized on a real time heat map. This project was selected in the top 20 out of 350 teams in Inmobi Hack Day.

Hands Free Cake Book
"Microsoft BootCamp" Microsoft, 2014

demo video | app

Published a Windows Store App which uses hand gesture recognition to allow the user to navigate through steps of baking a cake. This is done so that the user would not have to touch the laptop/tablet with his/her messy cake hands while baking. The app currently has 5,000 plus downloads from 120 different countries.

Web History Based Mobile App Recommender

"Inmobi Hack Day" InMobi, 2014 | read more about hack @ devpost

Developed an android app recommender which recommends an app based on the user's web browsing history and not based on the app's he has downloaded so far, like is currently done. Three layers of filtering are done to recommend the most useful app to a user based on his present interests(based on content of most visited sites). We get the content tags of a website from SimilarWeb API and then scrape the play store to recommend an app which is most similar to the content (most number of similar tags). Extensively used Java, JSoup for parsing, JDBC. This project was selected in the top 33 of Inmobi Hack Day.

Got this minimalistic website design from here