Abstract—In sharing capability also emphasises a need

Abstract—In modern times image sharing in social networkshave seen exponential increase. The instant sharing capabilityalso emphasises a need for better privacy systems. Unfortunatelyfor common users; privacy protection for image sharing hasbeen a challenging task due to the burden of legacy privacyassigning settings in images sharing sites. This necessitates a needfor automatic privacy assigning on images. With deep featuresextracted from images based on object detection; privacy canbe assigned to images.

Custom privacy recommendations can bemade on images before sharing them on social networks. Keywords—CNN, privacy, object detection.I. INTRODUCTION In modern times sharing images has become an integralpart of our daily life. The advent of smartphones and digitalcameras have resulted in exponential increase in imagescaptured. The instantly connected devices also facilitatedin quick sharing of these images.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

Unfortunately with largevolumes of instant sharing the privacy is often threatened;while sharing such special moments. And once something,such as a photo, is posted online, it becomes a permanentrecord, which may be used for purposes we never expect.It can lead to unwanted disclosure and privacy violations.

Malicious attackers can take advantage of these unnecessaryleaks to launch context-aware attacks or even impersonationattacks, as seen lately by the huge number of cases of privacyabuse and unwarranted access. Cloud photo services are also not the best while it comes toprivacy; they have hidden policies that gives large organiztionaccess to very privacy sensitive images. Additionally asecurity attack can compromise the users privacy. Currentsocial networking sites or cloud services do not assist usersin making privacy decisions for images that they share online.

In particular, privacy of online images is inherently asubjective matter, dependent on the image owners privacyattitude, awareness, and the overall context wherein the imageis to be posted. This creates a very unique challenge wherein we need to factor in so many paramters to determinethe cause of a particular privacy setting. These issues arekept in centre and an attempt is made to learn more aboutthe specific parameters that define privacy well. And also toeffectively retrieve them from images. Furthermore providingsome mathematical relevance to the privacy predictions thatare being made. II.

METHODS AND DISCUSSIONS Ashwini et. al1 investigates various methods for imageprivacy classification which are both visual and textualfeatures. And determine how they fare in image privacyclassification.

From the various identified features it combinessmallest set of features and measure how well they predictprivacy. The features range from basic to sophisticatedfeatures. The features include Scale Invariant FeatureTansform (SIFT), User tags, EDGE Direction Coherence,Sentiment analysis and use of RGB maps.

Here SIFT is a visual feature transform which extracts andcreates a bag of visual words. The bag of visual words vectoris a visual vocabulary. It represent the numbers of orginalSIFT feature vectors that are mapped onto correspondingvisual words. Sentiment Analysis is used where a visual sentimentontology which is used to detect sentiments. Its build onpsychological theories and web mining carrying conceptscalled Adjective Noun Pairs(ANP).

Some examples of ANPare Bad Food , Cute Dog , Beautiful Sky. For instance froma series of 1200 defined ANP concepts present in a particularimage; privacy can be predicted. Faces are likely to be considered private.

Hence identifyingfaces helps to predict privacy. This will be an effective featureand face detection usually use convolutional neural networks.The EDGE detection coherence is used to detect views ,landscapes , sky , sceneries e.t.c. These are considered to bemore public. EDGE Direction Coherence, which uses EDGEDirection Histograms to define the texture in an image.

Thefeature stores the number of edge pixels that are coherent andnon- coherent. A division of coherence and non-coherenceis established by a threshold on the size of the connectedcomponent edges. This eventually helps in effective detectionof sceneries. In online file sharing sites; the use of annotations ortags in the image is common. These tags are defined byusers themselves which is a very valuable asset for privacyprediction.

But the difficulty lies in obtaining them; as theyare not gauranteed to be available. 1 The above comparison results prove the superiority of usingTAGS over other techniques. And when considered as a pair;SIFT and TAGS stood out especially when dataset increases. The deep convolutional network rose to prominencefollowing the work of Krishvesky et. al1. They trained adeep convolutional neural network that classify 1.2 millionimages into 1000 distinct classes.

The convolutional neuralnetworks mimicks the biological process of neurons inlearning image features. It facilitated in finding patterns ofimages using building blocks of convolution layer , poolinglayer and fully connected layers. In convolving layer the input is an array of pixels offixed resolution. There are kernels or filters with definedsize; the kernel or filters are used in identifying the featurespresent in various images. The filter performs elementwisemultiplication on the input.

Then it slides or convolves aroundthe input image such that it is fully covered by the filter.Theparameter ‘strides’ define the distance between centers ofneighbouring kernels. There can be overlapping of the kernels. This is followed by applying nonlinearity functions; herea ReLU function is used. Then in pooling by method suchas max pooling we reduce the size of the output from thatlayer.

The process of convolution , pooling are repeatedfurther through the layers untill we reach the fully connectedlayer. The fully connected layer has 4096 neurons. As namesuggests they are all connected to each other each output aprobability score. This score is eventually used to classifyinto the required classes in the last layer. It also used techniques of dropout , data segmentation thatprevented overfitting. This work was seminal in the field;which to lead to pathbreaking contribution in computer vision.The architecture was backbone fot future convolutional neuralnetworks. The idea of image detection goes farther than image classi-fication.

In detection it requires localization of objects withinin an image. Ross et. al3 discuss about region proposing andhow objects can be effectively detected in an image. The objectdetection consist of three modules : Category independent region proposal Large CNN to extract features Multiclass SVM for classification The architecture has input images of predefined resolution. Itextracts around 2000 bottom up region proposals. The featuresare forward propogated and a mean subtracted RGB imagethrough 5 convolutional and 2 fully connected CNN. Thusacheiving very high detection accuracy than past ImageNetmodels.

The detection used concepts such as ground truth box, intersection over union. Intersection over Union was theevaluation metric used to measure the accuracy of an objectdetector on a particular dataset. It is an evaluation metric;where for a bounding box was predicted during detection.Now ground-truth bounding boxes were the hand labeledbounding boxes from the testing set that specify where in theimage object is present.

The system accuracy was measuredbased on overlapping of ground and predicted boxes to totalarea of union by both. The object detection acheived by 3 was considerablyslower. The factor that slowed down RCNN was it had aconstant number of 2000 squares. These were selected asregions always in every single images. Then all these regionshad to be passed through CNN and then classified. This issuewas solved in Fast RCNN. In Fast RCNN there is only asingle forward pass. The proposed object regions convolutionmap is used further to detect the objects.

The architecture takes an entire image as the input and withproposed object regions initially. The image is processed inwhole with several convolutional and max pooling layers toproduce a feature map. Then for each object proposed regionit further has a Region of Interest pooling layer extractinga fixed length feature vector from the feature map. Theseare directed onto a fully connected layer which gives the 2 bounding boxes for various object classes detected. Fast RCNN resulted in cutting down the training andtesting time when compared to RCNN; almost halving thetime it took for RCNN. Additionally it got better results interms of the accuracy of detection and provided many finetuning capabilities which can be used to improve the accuracyacross a wide range of datasets. Transfer learning is a technique described by Dai et.

al6where a system experience is used to train another learningsystem. In the case of neural networks; a particular networkis trained on some data. The learning system gains knowledgefrom this data, which is compiled as weights of the network.These weights can be extracted and then transferred to anyother neural network. Instead of training the other neuralnetwork from scratch, the learned features are transfered. Further fine tuning technique is applied in the sphere ofuser training. The pre-trained networks usually are trained onstandard datasets which may not perform well for solving allproblems.

Hence we need to deploy the model for furthertraining. This training will yield changes to the learnedweights and depends on the learning weights. A more advanced approach is of freezing certain layerswhile using others. Here the weights of initial layers of themodel frozen while we retrain only the higher layers. Theinitial layers primarily detect a general shape of the object;while higher layers facilitate in discrimination.

Jun et. al5 deals with privacy alignment of objectsdetected in an image. The algorithm discussed in helps toprecisely assign image level level privacy to the object classesin the images. Here they rely on semantic similarity measureon initially defining a bag of object classes. These form acluster which is useful in privacy alignment.

?(Xi,Xj) = 1000 ?(Xil,Xjl)1 Here for Xi and Xj are object classes for the lth image. Andif object class similarity is defined as ?(Xil,Xjl)=1;ifXil =Xjl Thus based on the similarity the clusters of similar objectsare generated and to which the common privacy setting isapplied. The privacy setting can be public , protected andprivate. An additional relevance score is defined; whichdefines the relevance between privacy setting of object classto the image privacy setting labeled.

?(Ci,t)= ?(Ci,t)?(Ci,P) Here ‘P’ denotes the clusters integrated privacy setting;while ?(Ci,t) denotes a subset of images in the cluster thatcontains object class Ci and has the privacy setting ‘t’. Finallyprivacy setting with the largest object-privacy relevance score is finally selected for the given object class CiIII. SOLUTION Manually assigning privacy settings to each image each timecan be cumbersome. To avoid a possible loss of users privacy,it has become critical to develop automated approaches thatcan accurately predict the privacy settings for shared images. The idea is to automatically detect the privacy-sensitiveobjects from the images being shared, recognise their classes,and identify their privacy settings.

Based on the detectionresults, our system would be able to warn the image ownerswhat objects in the images need to be protected. It can alsoprovide recommended privacy settings in an instance. IV. CONCLUSION The idea of automating privacy assigning was surveyed. Theuse of deep features in images and deep tags are revealed to bethe best parameters in deduction of privacy in an image.

Deepfeature extraction has been done widely using the techniqueof Convolutional Neural Networks. Various CNN architecturesfrom Imagenet to Fast RCNN were considered. The modernFast RCNN exhibits state of art result in detection while alsobeing much faster than other object detection methods sur-veyed. Coupling the former with a privacy alignment algorithmsurveyed can be used to build an automated privacy assigningsystem. The challenge will be to reduce computation overheadof using CNN’s and improve the prediction accuracy.

x

Hi!
I'm Mary!

Would you like to get a custom essay? How about receiving a customized one?

Check it out