Visual alerts are commonly used in video monitoring and surveillance systems to mark events, presumably making them
more salient to human observers. Surprisingly, the effectiveness of computer-generated alerts in improving human
performance has not been widely studied. To address this gap, we have developed a tool for simulating different alert
parameters in a realistic visual monitoring situation, and have measured human detection performance under conditions
that emulated different set-points in a surveillance algorithm. In the High-Sensitivity condition, the simulated alerts
identified 100% of the events with many false alarms. In the Lower-Sensitivity condition, the simulated alerts correctly
identified 70% of the targets, with fewer false alarms. In the control condition, no simulated alerts were provided. To
explore the effects of learning, subjects performed these tasks in three sessions, on separate days, in a counterbalanced,
within subject design. We explore these results within the context of cognitive models of human attention and learning.
We found that human observers were more likely to respond to events when marked by a visual alert. Learning played a
major role in the two alert conditions. In the first session, observers generated almost twice as many False Alarms as in
the No-Alert condition, as the observers responded pre-attentively to the computer-generated false alarms. However, this
rate dropped equally dramatically in later sessions, as observers learned to discount the false cues. Highest observer
Precision, Hits/(Hits + False Alarms), was achieved in the High Sensitivity condition, but only after training. The
successful evaluation of surveillance systems depends on understanding human attention and performance.
Social tagging is an emerging methodology that allows individual users to assign semantic keywords to content on the
web. Popular web services allow the community of users to search for content based on these user-defined tags. Tags
are typically attached to a whole entity such as a web page (e.g., del.icio.us), a video (e.g., YouTube), a product
description (e.g., Amazon) or a photograph (e.g., Flickr). However, finding specific information within a whole entity
can be a difficult, time-intensive process. This is especially true for content such as video, where the information sought
may be a small segment within a very long presentation. Moreover, the tags provided by a community of users may be
incorrect, conflicting, or incomplete when used as search terms.
In this paper we introduce a system that allows users to create "micro-tags," that is, semantic markers that are attached to
subsets of information. These micro-tags give the user the ability to direct attention to specific subsets within a larger
and more complex entity, and the set of micro-tags provides a more nuanced description of the full content. Also, when
these micro-tags are used as search terms, there is no need to do a serial search of the content, since micro-tags draw
attention to the semantic content of interest. This system also provides a mechanism that allows users in the community
to edit and delete each others' tags, using the community to refine and improve tag quality. We will also report on
empirical studies that demonstrate the value of micro-tagging and tag editing and explore the role micro-tags and tag
editing will play in future applications.
A substantial portion of the text available online is of a kind that tends to contain many typos and ungrammatical
abbreviations, e.g., emails, blogs, forums. It is therefore not surprising that, in such texts, one can carry out
information-hiding by the judicious injection of typos (broadly construed to include abbreviations and acronyms).
What is surprising is that, as this paper demonstrates, this form of embedding can be made quite resilient.
The resilience is achieved through the use of computationally asymmetric transformations (CAT for short):
Transformations that can be carried out inexpensively, yet reversing them requires much more extensive semantic
analyses (easy for humans to carry out, but hard to automate). An example of CAT is transformations that
consist of introducing typos that are ambiguous in that they have many possible corrections, making them harder
to automatically restore to their original form: When considering alternative typos, we prefer ones that are also
close to other vocabulary words. Such encodings do not materially degrade the text's meaning because, compared
to machines, humans are very good at disambiguation. We use typo confusion matrices and word level ambiguity
to carry out this kind of encoding. Unlike robust synonym substitution that also cleverly used ambiguity, the
task here is harder because typos are very conspicuous and an obvious target for the adversary (synonyms are
stealthy, typos are not). Our resilience does not depend on preventing the adversary from correcting without
damage: It only depends on a multiplicity of alternative corrections. In fact, even an adversary who has boldly
"corrected" all the typos by randomly choosing from the ambiguous alternatives has, on average, destroyed
around w/4 of our w-bit mark (and incurred a high cost in terms of the damage done to the meaning of the
Text data forms the largest bulk of digital data that people encounter and exchange daily. For this reason the potential usage of text data as a covert channel for secret communication is an imminent concern. Even though information hiding into natural language text has started to attract great interest, there has been no study on attacks against these applications. In this paper we examine the robustness of lexical steganography systems.In this paper we used a universal steganalysis method based on language models and support vector machines to differentiate sentences modified by a lexical steganography algorithm from unmodified sentences. The experimental accuracy of our method on classification of steganographically modified sentences was 84.9%. On classification of isolated sentences we obtained a high recall rate whereas the precision was low.
This paper gives an overview of the research and implementation challenges we encountered in building an end-to-end natural language processing based watermarking system. With natural language watermarking, we mean embedding the watermark into a text document, using the natural language components as the carrier, in such a way that the modifications are imperceptible to the readers and the embedded information is robust against possible attacks. Of particular interest is using the structure of the sentences in natural language text in order to insert the watermark. We evaluated the quality of the watermarked text using an objective evaluation metric, the BLEU score. BLEU scoring is commonly used in the statistical machine translation community. Our current system prototype achieves 0.45 BLEU score on a scale [0,1].
In this paper we discuss natural language watermarking, which uses the structure of the sentence constituents in natural language text
in order to insert a watermark. This approach is different from techniques, collectively referred to as "text watermarking," which embed information by modifying the appearance of text elements,
such as lines, words, or characters. We provide a survey of the current state of the art in natural language watermarking and introduce terminology, techniques, and tools for text processing. We also examine the parallels and differences of the two watermarking domains and outline how techniques from the image watermarking domain may be applicable to the natural language watermarking domain.