The StarSight Project

by S. P. Perone

  Heatedly, the Senator shared his suspicions and rationale with Carothers, as he had discussed with Ellen the previous night. When he had finished, Carothers looked thoughtfully at the Senator for several seconds, sitting back in his chair and raising his hand to stroke the stubble on his chin. Finally, he spoke. “Gerry, you’ve been working under a misconception for a long time. If the CIA had anything to do with getting the insider information to you, I don’t know anything about it. Yes. It’s possible…but very unlikely that some other branch would do this without my knowledge. And, that also means no one here could have leaked the information to Lee.”

  Staring blankly at Carothers, the Senator slowly slumped back in his chair, feeling some of the anger draining away. “Forgive me for my intensity this morning, Nathan. I’m not really angry with you. I’m angry with this asshole, Lee, and the way he blind-sided Ellen yesterday.”

  “Because we go back a long way,” the Senator continued, “I have to believe what you’re telling me, Nathan. But, that leaves one very important unanswered question: Who’s responsible for the insider information I received?”

  “Yes…and, there’s one more question, isn’t there, Gerry?What do they want from you? ”

  The Senator’s silence prodded Carothers to continue. “We’ve had Lee under surveillance for the past several months. He got our attention a few months ago when he made three brief trips to Switzerland within as many weeks. After the first two, we tracked him to see what was going on. And, we began working with Interpol to see what we could find out about the first two trips.

  “We learned that theSentinel was paying for these expensive trips. And, we found out who his contact was in Zurich. He’s a Saudi national, by the name of Muhammad Mustafa, currently on a Swiss work visa. He’s employed by the Sovereign Bank to provide liaison with several big Arab accounts. He is almost certainly the source of Lee’s information about your wife’s investments.

  “After beginning our surveillance, we learned about the story he was putting together. We know he contacted you; and we know about his visit to Ellen yesterday.”

  “Well, Goddammit, Nathan,” the Senator cried, rising to his feet, and looking down at his friend, “why didn’t you do something about it?”

  Unresponsive, Carothers simply returned the Senator’s hysterical outburst with a stern gaze that had the effect of directing him back into his seat. Then he answered the Senator’s question.

  “Trust me, Gerry,” Carothers began calmly, “Lee’s story will not be published. He wants to show a CIA connection to your campaign funding. He hasn’t found any evidence. There isn’t any. He’s been bluffing you…and Ellen…hoping to get an admission from one of you about CIA involvement. Fortunately, neither of you said anything. Ellen, because she truly knew nothing; and, you, because you were ‘stonewalling’.

  “Even though you’re feeling guilty for not telling Ellen about this situation before Lee showed up, you did the right thing. If you had shared your suspicions about the CIA, she might have been tricked into saying something Lee could use as confirmation for his story.”

  For a few moments, the two gentlemen sat back in their chairs, silent. Carothers waited for the Senator to voice the questions forming in his agile mind. It didn’t take long.

  “All right, Nathan,” he began, “who was responsible for providing me with the investment advice? And why? I know you’ve figured it out. Perhaps it’s time you shared that information with me.”

  Pondering whether this was indeed the right time, Carothers leaned forward, elbows on his desk, tented fingers hiding the lower half of his face. After retaining that pose for nearly a minute, and reaching a decision, he sat back once again, crossed his legs, lowered his hands, and began to speak, slowly and deliberately.

  “What I’m about to tell you now, Gerry, is mostly speculation. When you hear, you will know why I could not say anything earlier. Perhaps I shouldn’t be telling you now. In fact, you must agree not to take any independent action based on this information. Agreed?”

  A slow, solemn nod from the Senator permitted Carothers to continue.

  “We suspect that the al-Qa’eda terrorist organization was responsible for the campaign fund caper. Their motivation was to put you in a vulnerable position, which they could exploit at an opportune time. Considering the timing, it appears they had hoped to use this ploy to de-rail the StarSight project…possibly by discrediting the one person most responsible for its funding and progress.You .

  “Lee’s story should have hit the newsstands a couple months ago. The only reason it didn’t is because his contact in Zurich wasn’t able to fabricate credible evidence of CIA involvement. That’s where they miscalculated. They didn’t count on Lee or his Editor having the integrity to hold publication until they got credible confirmation. That forced Lee to harass you and Ellen, and to snoop around elsewhere.”

  “Why didn’t the terrorist group just give the story to some less scrupulous tabloid? There’s plenty of them around,” the Senator interrupted.

  “Don’t know. Perhaps they figured most people wouldn’t take one of those rags seriously,” Carothers offered.

  “Do you know anything about this contact in Zurich?” the Senator inquired.

  “This is where it begins to get very interesting,” Carothers replied. “We’ve been checking back to see where he’s been, who his friends are, what newspapers he reads, his lifestyle, his politics, etc. A few flags have popped up. First, when he lived in Saudi Arabia, Mustafa may have been connected to an al-Qa’eda cell, with links to bin Laden in Afghanistan.

  “Second, he travels all over Western Europe, Eastern Europe, and the Middle East as liaison to Arab clients of the Sovereign Bank. The one country where there’s no record of a visit, however, is Afghanistan. Sound fishy?”

  “Then, he has only one ‘client’ in the U.S. Actually, we don’t know that it is a ‘client’. We only know that there has been some contact between Mustafa and this very prominent Arab-American citizen. This gentleman sees Mustafa in Zurich, but never at the bank. They attend the same cocktail parties; go skiing at the same resorts; even run into each other at the same soccer matches.”

  “Sounds interesting,” the Senator interjected, his eyebrows arched.

  “More interesting is the identity of this prominent citizen…Ahmed Sharif,” Carothers stated, watching for the Senator’s reaction.

  Stunned at the mention of the man, whom he and Ellen considered to be their dear friend, the Senator’s jaw dropped as his eyes widened. “What? Ahmed knows Mustafa? There’s no way Ahmed is connected to terrorists! He’s as loyal to this country as Thomas Jefferson, for Christ’s sake! You must be mistaken.”

  Choosing to remain silent, Carothers calmly returned the Senator’s passionate gaze. He knew it was not necessary to debate this issue. Once the shock wore off, the Senator would digest and accept this new information about his friend, Sharif. Wryly, Carothers reflected on how jaded he himself had become. These kinds of revelations no longer surprised him.

  After nearly a full minute of silent fuming, his brain turning over the many morsels of information delivered by Carothers, the Senator re-focused his eyes on Carothers’ stoic face. “What does it all mean, Nathan?” he pleaded. “I feel like my world’s falling apart. Next, you’ll be telling me that I leaked the information about StarSight to the Goddam terrorists!”

  Carothers’ continued silence, remaining motionless in his chair, while calmly returning the Senator’s gaze, was more damning than anything that could have been said. Slowly, the Senator slumped back in his chair, tossing his head back against the high padded back, looking up at the pale high ceiling of Carothers’ office, and breathing a barely audible sigh. The full weight of the world had descended upon him with a velocity that rocked his very foundation. “My God. What have I done, Nathan?” he whispered. “What have I done?”

  “You haven’t done anything, old friend,” Carothers responded gently. “There were many possibl
e sources of StarSight information accessible to al-Qa’eda…or any other terrorist group. Theymay have found out about the project from your office…but it really doesn’t matter. Not now. Spies extract intelligence from us all the time, without our knowledge, sometimes despite heroic attempts to thwart their efforts. In this case, I think we may have been careless. Administering this project through DOE and using university scientists to conduct the research left too many insecure access points.

  “What’s important now, Gerry, is what we do with the information we have. We have an opportunity to take some action. We’re certain they don’t know we’re aware of the Mustafa and Sharif connection. We need to take advantage of this now. And…Gerry…I need your help.”

  These last words sparked a hint of reaction from beneath the dull glaze that had descended over the Senator’s eyes. “My help?” he murmured. “What can I do?”

  Chapter 16

  Successful Predictions?

  The first heavy rain of the winter season had spread over the entire Bay Area. The cleansing of the air produced a sweet moist smell, enhanced by the pungent aroma of eucalyptus oil, rain-distilled from the leaves of large numbers of colorful eucalyptus trees lining many of the country roads winding through the Livermore Valley. Although the hills were still a summer-baked golden brown, the promise of a lush green cloak was not far away.

  The four scientists, huddled together in front of a computer screen in Building 451 of the Lawrence Livermore Lab, were completely oblivious to the weather change that had rolled through the valley since they had begun working at six o’clock that morning. Like many at the Lab, the ASCI White supercomputer facility had no windows to the outside. Located at nearly the exact center of the Lab site, the facility had a rigidly controlled environment so that ASCI White could achieve its world-class speed of 12.7 trillion operations per second.

  Painfully aware that the beginning of the holiday season was only a week away, Tony Shane and his colleagues were examining the latest output of the neural network clustering program. Because she had slowly taken over the lead effort to interpret the cluster analysis results, Sarah was explaining to Tony, Barry, and Sharon the ostensibly random scattering of hundreds of symbols on the computer screen.

  “Each point represents a date in 1995 to 1998,” she began, “and its color and position on the screen is computed by the StarSight neural network. The blue symbols represent dates where the network predicted no terrorist events anywhere in the world. The red, orange, yellow, green, pink, and purple symbols represent dates where specific types of terrorist events were predicted: explosions, suicide bombings, hijackings, abductions, cyberspace attacks, and miscellaneous events that haven’t been categorized. The input data for each point are those for the second week preceding each date. The data we used to train this network were obtained from 1985 to 1994.

  “Except for the purple symbols, the different-colored points are nearly completely separated into distinct clusters. The cleanly separated clusters represent predictions that were accurate. This means we were able to predict the date and type of terrorist event that occurred, with about a seven-day advance warning, for this test set of data. Accuracy over 95 percent. False positives under 5 percent.

  “This is about where we were last week. ‘What?’ and ‘when?’ But, now, when we re-organize the inputs geographically, we can use supervised pattern recognition to ask the question: Where?”

  As the rest of the group watched, Sarah used the workstation keyboard to call up the next mapping on the computer screen. This time, there were several sparsely populated clusters distributed around the screen. Each cluster contained symbols of a different color. Included in many of these clusters were points represented by different letters of the alphabet. For example, the cluster of green symbols contained three points represented by three “A’s” on the screen. The blue symbol cluster contained two “B’s”, and so on.

  “What you’re looking at on the screen,” Sarah explained, “are the results for only a handful of countries recording more than ten terrorist events in the 1985 to 1994 decade. The different-colored clusters represent those countries. Each letter symbol on the screen represents an event predicted by the test data (1995 to 1998), and the cluster it’s found in tells you which country will experience the event. The results are 100% accurate.”

  “Are these all similar events, Sarah,” Nagle asked. “Or do you have to train the network separately for each type of event?”

  “Good question, Barry,” she replied. “These results are for bombings only. That’s a big limitation. Sharon, Bill, and Anna have been looking at an all-inclusive training set, and Sharon can fill us in. But, at this point, we can pinpoint the country as long as we train the network with a particular kind of event.”

  “But, the first step in the prediction step will tell you what type of event would be involved, right?” Shane interjected.

  “That’s true. And, if Sharon’s approach works, we will be able to identify the country in the second step, regardless of the type of event.”

  Turning to Sharon Carson, Shane inquired, “Can you describe your results to us, Sharon? Or is it too soon?”

  “Yes. It is too soon,” Sharon replied. “But I can give you a progress report. Sarah came up with a way of condensing satellite images so that a single 64-bit code captures all the useful information in each image. What I’m working on now is using the same principles to condense the demographics data inputs also. The idea is that, we can input more ‘information’ to the neural network, with reduced complexity…which usually gives us greater reliability. I can show you the results we’ve obtained so far.”

  Replacing Sarah at the workstation keyboard, Sharon used several keystrokes to bring up a new image. Again, several clusters, each containing different-colored symbols, appeared on the screen. In contrast to the preceding images, however, these clusters were not separated, but were overlapped. Only the use of different colored symbols told the observer that more than a single blob covered the screen. No letter symbols appeared on the screen.

  Before Sharon could speak, Shane commented on the display before them. “I don’t see the test data predictions here. I assume that’s because you couldn’t assign a geographical location with the severe cluster overlaps. But, can’t you get around that with supervised pattern recognition?”

  “Let me explain, Tony. We get very good supervised training results with neural networks. You can see 100% separation of the clusters. But, when you try to predict on ‘future’ data, you just get nonsense.”

  “How do you explain that?”

  “It’s because there aren’t enough historical events in the training data for each geographical location. For neural networks, with this kind of deficient training set, they’re able to ‘learn’ very complex discrimination rules…which work perfectly for the training set…but are not truly ‘general’. When the trained network is applied to new data, it makes unreliable decisions.

  “It’s like, learning how to find a rest room at Yankee Stadium doesn’t teach you how to find one in the LA Coliseum. But, if you visit enough stadiums, you learn ageneral rule…like, rest rooms are always placed between beer and hot dog concessions…which works most of the time, even on your first trip to Wrigley Field.

  “What I’m saying is that a neural network could learn rules that work perfectly to find a rest room in any stadium it’s ever seen…but these rules would not be ‘general’, if it’s only seen a few different stadiums. So they fall apart when applied to the first new stadium.

  “That’s the problem here. In our case, the neural network gets ‘overtrained’ on deficient data.”

  “What if we deliberatelyundertrain the network?” Shane interjected. “Just stop the training at less than 100% recognition of the training set? Wouldn’t that produce a more general solution?”

  “That’s a good idea, Tony,” Sarah inserted, “but choosing an arbitrary stopping point for training is a very tricky procedure.�
�

  “What we really need to do,” she continued, “is modify the data so that it isn’t deficient in one or more of the classes of events.”

  “How can we possibly do that, Sarah?” Sharon asked. “We can’tfabricate terrorist events.”

  “Yes, we can…in a way,” Sarah insisted. “What if we have a unique event in the training set? One of a kind…like that crude attempt in 1989 to disable the crew of one of our naval destroyers in the Mediterranean by contaminating the water supply with a time-released bacteria. What we can do is artificially insert this same event into our database at several different times.”

  “How in the world could you do that?” Tony inquired. “You would have to overlay all of the demographic and satellite data that preceded the real event into the rest of the database for every ‘simulated’ event. Wouldn’t that mess up your ability to detect other real events?”

  “You’re right, Tony,” Sarah responded. “But, I’ve been thinking of a way to get around that problem. You know how we have to ‘normalize’ the data. We subtract the average value of each type of input from each data point going into the neural network? We do that so the neural network is not influenced by the background variations in the data…those that are always present. Only the variations related to erratic events…like the terrorist attacks…affect the network. If any of these ‘solitary’ terrorist events occurred in the middle of a one-month interval when no other terrorist event occurred…then we can take the normalized data for two weeks prior to that event, and add it to the existing data anywhere else in the database. That way we can simulate the same type of event.”

  “Wait a minute, Sarah,” Nagle interrupted. “How would younot distort the data set byadding simulated data directly to your data set?”

  “Keep in mind, Barry,” Sarah replied, “that we’re addingnormalized data. Tell me what would happen if we selected a two-week time interval from our historical database where no terrorist event occurred? Where only normal variations in the data would occur?”